@datakable/core 2.0.2 → 2.1.0

package/dist/features/search/search-index.cjs

@@ -34,7 +34,7 @@ var import_row = require("../../pipeline/row.cjs");
 var import_capability = require("../../versioning/capability.cjs");
 var CHUNK = 4096;
 var CELLS = 16384;
-var hasExSearch = () => [import_ex_search.ExSearch.prototype.search, import_ex_search.ExSearch.prototype.setData, import_ex_search.Scorer.fuzzy].every(
+var hasExSearch = () => [import_ex_search.ExSearch.prototype.search, import_ex_search.ExSearch.prototype.setData, import_ex_search.ExSearch.prototype.updateConfig].every(
   (f) => typeof f === "function"
 );
 function probeExSearch(emit) {
@@ -43,8 +43,8 @@ function probeExSearch(emit) {
     {
       feature: "search",
       dependency: "ex-search",
-      capability: "ExSearch, Scorer.fuzzy",
-      supported: "^1.1.1",
+      capability: "ExSearch (fuzzyMode: 'word')",
+      supported: "^1.2.0",
       fallback: "case-insensitive exact / prefix / substring matching (no fuzzy matching)"
     },
     emit
@@ -129,7 +129,7 @@ function* buildIndex(ctx, config, useExSearch) {
   const fold = (v) => config.caseSensitive ? v : v.toLowerCase();
   const foldQuery = (q) => fold(q.normalize("NFC"));
   const at = (k, r) => fold(reads[k](r));
-  return useExSearch ? exSearchIndex(ctx.data.length, keys, at, foldQuery, config) : containsIndex(keys, at, foldQuery, config);
+  return useExSearch ? exSearchIndex(keys, at, foldQuery, config) : containsIndex(keys, at, foldQuery, config);
 }
 function* chunked(rows, columns, matchChunk) {
   const size = Math.max(256, Math.floor(CELLS / columns));
@@ -141,33 +141,30 @@ function* chunked(rows, columns, matchChunk) {
   }
   return rankHits(hits);
 }
-function projection(size, keys, at) {
-  let rows;
+function projection(keys, at) {
+  let rows = new Uint32Array(0);
   const target = [];
-  const length = () => rows ? rows.length : size;
-  const item = (j) => {
-    const r = rows ? rows[j] : j;
-    const o = { i: r };
-    for (let k = 0; k < keys.length; k++) o[keys[k].name] = at(k, r);
-    return o;
-  };
   const items = new Proxy(target, {
     get(t, p) {
-      if (p === "length") return length();
-      const j = (0, import_row.arrayIndex)(p, length());
-      return j >= 0 ? item(j) : t[p];
+      if (p === "length") return rows.length;
+      const j = (0, import_row.arrayIndex)(p, rows.length);
+      if (j < 0) return t[p];
+      const r = rows[j];
+      const o = { i: r };
+      for (let k = 0; k < keys.length; k++) o[keys[k].name] = at(k, r);
+      return o;
     },
-    has: (t, p) => (0, import_row.arrayIndex)(p, length()) >= 0 || p in t
+    has: (t, p) => (0, import_row.arrayIndex)(p, rows.length) >= 0 || p in t
   });
   return {
     items,
-    /** Shows `chunk` (row indices), or every row again (`undefined`). */
+    /** Shows `chunk` (row indices). */
     show(chunk) {
       rows = chunk;
     }
   };
 }
-function exSearchIndex(size, keys, at, fold, config) {
+function exSearchIndex(keys, at, fold, config) {
   const engine = new import_ex_search.ExSearch({
     keys,
     ...config,
@@ -176,36 +173,18 @@ function exSearchIndex(size, keys, at, fold, config) {
     // stable and avoids radix's score quantisation.
     sortAlgorithm: "tim"
   });
-  const max = config.fuzzyMaxDistance;
-  let q = "";
-  const near = (k, r) => {
-    const v = at(k, r);
-    if (v.includes(q) || /\s/.test(q)) return v;
-    let best = "";
-    let top = 0;
-    const words = v.split(/\s+/);
-    for (const w of words[1] ? [v, ...words] : words) {
-      const s = Math.abs(w.length - q.length) > max ? 0 : import_ex_search.Scorer.fuzzy(w, q, max, true);
-      if (s > top) {
-        best = w;
-        top = s;
-      }
-    }
-    return best;
-  };
-  const view = projection(size, keys, near);
+  const view = projection(keys, at);
   engine.setData(view.items);
   return {
     match(rows, query) {
-      const folded = fold(query);
+      const q = fold(query);
+      const fuzzyMode = /\s/.test(q) ? "value" : "word";
       return chunked(rows, keys.length, (chunk) => {
-        q = folded;
+        engine.updateConfig({ fuzzyMode });
         view.show(chunk);
-        try {
-          return engine.search(q);
-        } finally {
-          view.show(void 0);
-        }
+        const hits = engine.search(q);
+        engine.updateConfig({});
+        return hits;
       });
     }
   };

package/dist/features/search/search-index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../src/features/search/search-index.ts"],"sourcesContent":["import { ExSearch, Scorer } from 'ex-search'\nimport { getValue } from '../../column/get-value'\nimport '../../diagnostics/process-guard'\nimport type { Emit } from '../../diagnostics/types'\nimport type { RowModelContext } from '../../feature/types'\nimport { nullishOf, type SlicedFormatting } from '../../pipeline/column-cache'\nimport { resolveColumns } from '../../pipeline/columns'\nimport { arrayIndex } from '../../pipeline/row'\nimport { detectCapability, hasCapability } from '../../versioning/capability'\n\n/** Matcher settings (the client options of `search()`, defaults applied). */\nexport interface MatchConfig {\n  readonly threshold: number\n  readonly fuzzyMaxDistance: number\n  readonly caseSensitive: boolean\n}\n\n/**\n * One projected row as ex-search reads it: `i` is the row index (results map back by it, never by\n * identity: ex-search returns copies), `c0`, `c1`, … the formatted values of the searchable\n * columns. Flat keys also sidestep ex-search's top-level-key limitation for nested paths. Items\n * are built per chunk while matching and dropped after it (§9.3 memory): none is kept.\n */\nexport type Item = Record<string, string | number> & { readonly i: number }\n\n/** A match: the row index and its relevance (0–1, weight applied). */\nexport interface Hit {\n  readonly i: number\n  readonly score: number\n}\n\n/** The matcher of one (data, columns) pair: references to column values, no per-row copy. */\nexport interface SearchIndex {\n  /**\n   * A time-sliced job (§9.1) matching `query` (trimmed, non-blank) against the rows `rows` (row\n   * indices), in chunks: the hits in relevance order ({@link rankHits}).\n   */\n  match(rows: Uint32Array, query: string): Generator<undefined, Uint32Array, undefined>\n}\n\n/** Rows per chunk between two `yield`s of a scan (one value per row). */\nconst CHUNK = 4096\n\n/**\n * Cells (rows × searchable columns) matched per chunk between two `yield`s (#40): a table with\n * many searchable columns matches fewer rows per chunk, so a chunk stays a few milliseconds.\n */\nconst CELLS = 16_384\n\ninterface Key {\n  readonly name: string\n  readonly weight: number\n}\n\n/** The formatted value of row `r` of one searchable column. */\ntype Read = (r: number) => string\n\n// A throw (no `ExSearch`) counts as missing (hasCapability).\nconst hasExSearch = () =>\n  [ExSearch.prototype.search, ExSearch.prototype.setData, Scorer.fuzzy].every(\n    (f) => typeof f === 'function',\n  )\n\n/** Feature-detects ex-search once (§8.5); `false` → the contains fallback. */\nexport function probeExSearch(emit: Emit): boolean {\n  // DK_DEP_INCOMPATIBLE is a development check (§8.5); the fallback is not.\n  return (process.env.NODE_ENV !== \"production\")\n    ? detectCapability(\n        hasExSearch,\n        {\n          feature: 'search',\n          dependency: 'ex-search',\n          capability: 'ExSearch, Scorer.fuzzy',\n          supported: '^1.1.1',\n          fallback: 'case-insensitive exact / prefix / substring matching (no fuzzy matching)',\n        },\n        emit,\n      )\n    : hasCapability(hasExSearch)\n}\n\n/**\n * A numeric column's formatted value without cached strings (#27): what the default formatter\n * shows, `String(v)`, or `''` for a `null` / `undefined` cell (stored as `NaN`, like a real `NaN`,\n * which shows `'NaN'`).\n */\nfunction numberReader(values: Float64Array): Read {\n  const nullish = nullishOf(values)\n  return (r) => {\n    const v = values[r] as number\n    return !Number.isNaN(v) ? String(v) : nullish?.[r] ? '' : 'NaN'\n  }\n}\n\n/**\n * One pass over a column's values (`get(r)` per row, time-sliced by `CHUNK` rows): their text as\n * the default formatter shows a primitive (`String(v)`, `''` when nullish), read while matching\n * (#27: no formatted copy per row). `false` at the first object or function (a `Date`, an array):\n * the column cache formats those. `undefined` when a read throws. The values are read in NFC (#30),\n * normalised here, never per keystroke: the rows whose value is not NFC (rare: NFD input from\n * macOS, Thai or Vietnamese keyboards) keep their NFC form in a sparse map. Only values with a code\n * unit from U+0300 (combining marks and up) can change.\n */\nfunction* scan(\n  size: number,\n  get: (r: number) => unknown,\n): Generator<undefined, Read | false | undefined, undefined> {\n  const over = new Map<number, string>()\n  try {\n    for (let r = 0; r < size; r++) {\n      const v = get(r)\n      if (v !== null && (typeof v === 'object' || typeof v === 'function')) return false\n      if (typeof v === 'string' && /[\\u0300-\\uffff]/.test(v)) {\n        const n = v.normalize('NFC')\n        if (n !== v) over.set(r, n)\n      }\n      if (r % CHUNK === CHUNK - 1) yield\n    }\n  } catch {\n    return undefined\n  }\n  const read: Read = (r) => {\n    const v = get(r)\n    return v === null || v === undefined ? '' : String(v)\n  }\n  return over.size ? (r) => over.get(r) ?? read(r) : read\n}\n\n/**\n * Builds the matcher over the formatted values of the searchable columns (what the user sees), once\n * per data + columns, as a time-sliced job (§9.1): one column read per slice. Searchable columns are\n * those with a `search` weight > 0. When no column declares `search`, the columns of primitives\n * (strings, numbers, booleans, nullable) are searchable at weight 1, and a column of dates, arrays\n * or other objects is not (#27: it would keep a formatted string per row; it opts in with a\n * weight; `DK_SEARCH_COLUMN_SKIPPED` in development). `null` when nothing is searchable. Nothing\n * is copied per row (§9.3 memory): a key column of primitives without `format` is read from the\n * rows and stringified while matching (no values array, #27); an accessor column of primitives\n * reads the column cache's values the same way; only a column with `format` or an opted-in object\n * column (dates, arrays) keeps its formatted strings in the column cache. A key column whose read\n * throws goes through the column cache too, which reports it. The case is folded per chunk while\n * matching.\n */\nexport function* buildIndex(\n  ctx: RowModelContext,\n  config: MatchConfig,\n  useExSearch: boolean,\n): Generator<undefined, SearchIndex | null, undefined> {\n  const cols = resolveColumns(ctx.columns)\n  const declared = cols.list.some((c) => (c.def as { search?: unknown }).search !== undefined)\n  const data = ctx.data\n  const n = data.length\n  const keys: Key[] = []\n  const reads: Read[] = []\n  for (const col of cols.list) {\n    if (cols.byId.get(col.id) !== col) continue\n    const w = (col.def as { search?: unknown }).search\n    const weight = declared ? (typeof w === 'number' && w > 0 ? w : 0) : 1\n    if (weight === 0) continue\n    const path = col.format ? undefined : col.path\n    let failed = false\n    const onError = () => {\n      failed = true\n    }\n    let read = path ? yield* scan(n, (r) => getValue(data[r], path, onError)) : undefined\n    if (failed || (read === false && declared)) read = undefined\n    if (read === undefined) {\n      const values = ctx.getColumnValues(col.id)\n      if (!values) continue\n      yield // reading a column is a task of its own (#40)\n      const numbers = values instanceof Float64Array\n      if (numbers && !col.format) read = numberReader(values)\n      else {\n        // Zero-config: a column of objects is not searched, `format` or not (#27).\n        const own =\n          numbers || (declared && col.format) ? undefined : yield* scan(n, (r) => values[r])\n        if (own === false && !declared) read = false\n        else if (own && !col.format) read = own\n        else {\n          // Time-sliced in a table run (#40): dates or a heavy `format` are not one long task.\n          const sliced = (ctx as RowModelContext & SlicedFormatting).formattedSliced\n          const formatted = sliced ? yield* sliced(col.id) : ctx.getFormattedValues(col.id)\n          if (!formatted) continue\n          read = yield* scan(n, (r) => formatted[r])\n        }\n      }\n    }\n    if (!read) {\n      if ((process.env.NODE_ENV !== \"production\") && read === false) {\n        ctx.emit('DK_SEARCH_COLUMN_SKIPPED', { column: col.id, index: col.index })\n      }\n      continue\n    }\n    keys.push({ name: `c${keys.length}`, weight })\n    reads.push(read)\n    yield\n  }\n  if (keys.length === 0) return null\n  // Values (NFC already) are folded per chunk while matching; the query is normalised as well, so\n  // NFC and NFD spellings match each other like in the `contains` filter (#30).\n  const fold = (v: string) => (config.caseSensitive ? v : v.toLowerCase())\n  const foldQuery = (q: string) => fold(q.normalize('NFC'))\n  const at = (k: number, r: number) => fold((reads[k] as Read)(r))\n  return useExSearch\n    ? exSearchIndex(ctx.data.length, keys, at, foldQuery, config)\n    : containsIndex(keys, at, foldQuery, config)\n}\n\n/** Matches `rows` chunk by chunk (`score` decides alone, so chunks match like the whole). */\nfunction* chunked(\n  rows: Uint32Array,\n  columns: number,\n  matchChunk: (chunk: Uint32Array) => Hit[],\n): Generator<undefined, Uint32Array, undefined> {\n  const size = Math.max(256, Math.floor(CELLS / columns))\n  if (rows.length <= size) return rankHits(matchChunk(rows))\n  const hits: Hit[] = []\n  for (let k = 0; k < rows.length; k += size) {\n    for (const h of matchChunk(rows.subarray(k, k + size))) hits.push(h)\n    yield\n  }\n  return rankHits(hits)\n}\n\n/**\n * The projection as ex-search reads it: an array (iteration, `length`, index reads) of `{ i, c0, … }`\n * items, each built when read and not kept. It shows every row of `data` (`size`), or one chunk of\n * rows while that chunk is matched; ex-search holds this one view (`setData` once).\n */\nfunction projection(size: number, keys: readonly Key[], at: (k: number, r: number) => string) {\n  let rows: Uint32Array | undefined\n  // Never written, and its own `length` stays 0: `length` is read through the trap below (setting\n  // a large `length` would allocate a backing store).\n  const target: Item[] = []\n  const length = () => (rows ? rows.length : size)\n  const item = (j: number): Item => {\n    const r = rows ? (rows[j] as number) : j\n    const o = { i: r } as Record<string, string | number>\n    for (let k = 0; k < keys.length; k++) o[(keys[k] as Key).name] = at(k, r)\n    return o as Item\n  }\n  const items = new Proxy(target, {\n    get(t, p) {\n      if (p === 'length') return length()\n      const j = arrayIndex(p, length())\n      return j >= 0 ? item(j) : t[p as keyof Item[]]\n    },\n    has: (t, p) => arrayIndex(p, length()) >= 0 || p in t,\n  })\n  return {\n    items,\n    /** Shows `chunk` (row indices), or every row again (`undefined`). */\n    show(chunk: Uint32Array | undefined) {\n      rows = chunk\n    },\n  }\n}\n\nfunction exSearchIndex(\n  size: number,\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  // Values and query are folded already (buildIndex).\n  const engine = new ExSearch<Item>({\n    keys: keys as Key[],\n    ...config,\n    caseSensitive: true,\n    // DataKable re-ranks with exact scores and an index tie-break; tim keeps ex-search's own pass\n    // stable and avoids radix's score quantisation.\n    sortAlgorithm: 'tim',\n  })\n  // ex-search measures the edit distance to the whole value, so a typo of one word of a multi-word\n  // cell never matched (#10). For a one-word query, a value without it is shown as the best fuzzy\n  // match of itself and its words, or as '' (scores 0, as the value would). A value with the query\n  // stays whole for the exact / prefix / substring scores. Only strings within `max` of the query's\n  // length are measured (an edit distance is at least the length difference).\n  const max = config.fuzzyMaxDistance\n  let q = ''\n  const near = (k: number, r: number) => {\n    const v = at(k, r)\n    if (v.includes(q) || /\\s/.test(q)) return v\n    let best = ''\n    let top = 0\n    const words = v.split(/\\s+/)\n    for (const w of words[1] ? [v, ...words] : words) {\n      const s = Math.abs(w.length - q.length) > max ? 0 : Scorer.fuzzy(w, q, max, true)\n      if (s > top) {\n        best = w\n        top = s\n      }\n    }\n    return best\n  }\n  const view = projection(size, keys, near)\n  engine.setData(view.items)\n  return {\n    match(rows, query) {\n      const folded = fold(query)\n      return chunked(rows, keys.length, (chunk) => {\n        // Set per chunk, like the view: another job may have matched in between.\n        q = folded\n        view.show(chunk)\n        try {\n          // Results are copies of the items (`Omit<Item, …>` drops `i` from the type, not the value).\n          return engine.search(q) as unknown as Hit[]\n        } finally {\n          view.show(undefined)\n        }\n      })\n    },\n  }\n}\n\n/** The fallback when ex-search is missing or incompatible: exact 1 · prefix 0.8 · substring 0.65. */\nfunction containsIndex(\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  return {\n    match(rows, query) {\n      const q = fold(query)\n      return chunked(rows, keys.length, (chunk) => {\n        const hits: Hit[] = []\n        for (let j = 0; j < chunk.length; j++) {\n          const r = chunk[j] as number\n          let best = 0\n          for (let k = 0; k < keys.length; k++) {\n            const v = at(k, r)\n            const raw = v === q ? 1 : v.startsWith(q) ? 0.8 : v.includes(q) ? 0.65 : 0\n            const w = (keys[k] as Key).weight\n            if (raw * w > best) best = raw * w\n          }\n          if (best > 0 && best >= config.threshold) hits.push({ i: r, score: best })\n        }\n        return hits\n      })\n    },\n  }\n}\n\n/**\n * Hits in relevance order (§6.3): score descending, ties by original row index. Enforced here so\n * the order never depends on the matcher's own ranking. Sorts `hits` (a fresh array) in place.\n */\nexport function rankHits(hits: Hit[]): Uint32Array {\n  return Uint32Array.from(\n    hits.sort((a, b) => b.score - a.score || a.i - b.i),\n    (h) => h.i,\n  )\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,uBAAiC;AACjC,uBAAyB;AACzB,2BAAO;AAGP,0BAAiD;AACjD,qBAA+B;AAC/B,iBAA2B;AAC3B,wBAAgD;AAiChD,IAAM,QAAQ;AAMd,IAAM,QAAQ;AAWd,IAAM,cAAc,MAClB,CAAC,0BAAS,UAAU,QAAQ,0BAAS,UAAU,SAAS,wBAAO,KAAK,EAAE;AAAA,EACpE,CAAC,MAAM,OAAO,MAAM;AACtB;AAGK,SAAS,cAAc,MAAqB;AAEjD,SAAQ,QAAQ,IAAI,aAAa,mBAC7B;AAAA,IACE;AAAA,IACA;AAAA,MACE,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,MACX,UAAU;AAAA,IACZ;AAAA,IACA;AAAA,EACF,QACA,iCAAc,WAAW;AAC/B;AAOA,SAAS,aAAa,QAA4B;AAChD,QAAM,cAAU,+BAAU,MAAM;AAChC,SAAO,CAAC,MAAM;AACZ,UAAM,IAAI,OAAO,CAAC;AAClB,WAAO,CAAC,OAAO,MAAM,CAAC,IAAI,OAAO,CAAC,IAAI,UAAU,CAAC,IAAI,KAAK;AAAA,EAC5D;AACF;AAWA,UAAU,KACR,MACA,KAC2D;AAC3D,QAAM,OAAO,oBAAI,IAAoB;AACrC,MAAI;AACF,aAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,YAAM,IAAI,IAAI,CAAC;AACf,UAAI,MAAM,SAAS,OAAO,MAAM,YAAY,OAAO,MAAM,YAAa,QAAO;AAC7E,UAAI,OAAO,MAAM,YAAY,kBAAkB,KAAK,CAAC,GAAG;AACtD,cAAM,IAAI,EAAE,UAAU,KAAK;AAC3B,YAAI,MAAM,EAAG,MAAK,IAAI,GAAG,CAAC;AAAA,MAC5B;AACA,UAAI,IAAI,UAAU,QAAQ,EAAG;AAAA,IAC/B;AAAA,EACF,QAAQ;AACN,WAAO;AAAA,EACT;AACA,QAAM,OAAa,CAAC,MAAM;AACxB,UAAM,IAAI,IAAI,CAAC;AACf,WAAO,MAAM,QAAQ,MAAM,SAAY,KAAK,OAAO,CAAC;AAAA,EACtD;AACA,SAAO,KAAK,OAAO,CAAC,MAAM,KAAK,IAAI,CAAC,KAAK,KAAK,CAAC,IAAI;AACrD;AAgBO,UAAU,WACf,KACA,QACA,aACqD;AACrD,QAAM,WAAO,+BAAe,IAAI,OAAO;AACvC,QAAM,WAAW,KAAK,KAAK,KAAK,CAAC,MAAO,EAAE,IAA6B,WAAW,MAAS;AAC3F,QAAM,OAAO,IAAI;AACjB,QAAM,IAAI,KAAK;AACf,QAAM,OAAc,CAAC;AACrB,QAAM,QAAgB,CAAC;AACvB,aAAW,OAAO,KAAK,MAAM;AAC3B,QAAI,KAAK,KAAK,IAAI,IAAI,EAAE,MAAM,IAAK;AACnC,UAAM,IAAK,IAAI,IAA6B;AAC5C,UAAM,SAAS,WAAY,OAAO,MAAM,YAAY,IAAI,IAAI,IAAI,IAAK;AACrE,QAAI,WAAW,EAAG;AAClB,UAAM,OAAO,IAAI,SAAS,SAAY,IAAI;AAC1C,QAAI,SAAS;AACb,UAAM,UAAU,MAAM;AACpB,eAAS;AAAA,IACX;AACA,QAAI,OAAO,OAAO,OAAO,KAAK,GAAG,CAAC,UAAM,2BAAS,KAAK,CAAC,GAAG,MAAM,OAAO,CAAC,IAAI;AAC5E,QAAI,UAAW,SAAS,SAAS,SAAW,QAAO;AACnD,QAAI,SAAS,QAAW;AACtB,YAAM,SAAS,IAAI,gBAAgB,IAAI,EAAE;AACzC,UAAI,CAAC,OAAQ;AACb;AACA,YAAM,UAAU,kBAAkB;AAClC,UAAI,WAAW,CAAC,IAAI,OAAQ,QAAO,aAAa,MAAM;AAAA,WACjD;AAEH,cAAM,MACJ,WAAY,YAAY,IAAI,SAAU,SAAY,OAAO,KAAK,GAAG,CAAC,MAAM,OAAO,CAAC,CAAC;AACnF,YAAI,QAAQ,SAAS,CAAC,SAAU,QAAO;AAAA,iBAC9B,OAAO,CAAC,IAAI,OAAQ,QAAO;AAAA,aAC/B;AAEH,gBAAM,SAAU,IAA2C;AAC3D,gBAAM,YAAY,SAAS,OAAO,OAAO,IAAI,EAAE,IAAI,IAAI,mBAAmB,IAAI,EAAE;AAChF,cAAI,CAAC,UAAW;AAChB,iBAAO,OAAO,KAAK,GAAG,CAAC,MAAM,UAAU,CAAC,CAAC;AAAA,QAC3C;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM;AACT,UAAK,QAAQ,IAAI,aAAa,gBAAiB,SAAS,OAAO;AAC7D,YAAI,KAAK,4BAA4B,EAAE,QAAQ,IAAI,IAAI,OAAO,IAAI,MAAM,CAAC;AAAA,MAC3E;AACA;AAAA,IACF;AACA,SAAK,KAAK,EAAE,MAAM,IAAI,KAAK,MAAM,IAAI,OAAO,CAAC;AAC7C,UAAM,KAAK,IAAI;AACf;AAAA,EACF;AACA,MAAI,KAAK,WAAW,EAAG,QAAO;AAG9B,QAAM,OAAO,CAAC,MAAe,OAAO,gBAAgB,IAAI,EAAE,YAAY;AACtE,QAAM,YAAY,CAAC,MAAc,KAAK,EAAE,UAAU,KAAK,CAAC;AACxD,QAAM,KAAK,CAAC,GAAW,MAAc,KAAM,MAAM,CAAC,EAAW,CAAC,CAAC;AAC/D,SAAO,cACH,cAAc,IAAI,KAAK,QAAQ,MAAM,IAAI,WAAW,MAAM,IAC1D,cAAc,MAAM,IAAI,WAAW,MAAM;AAC/C;AAGA,UAAU,QACR,MACA,SACA,YAC8C;AAC9C,QAAM,OAAO,KAAK,IAAI,KAAK,KAAK,MAAM,QAAQ,OAAO,CAAC;AACtD,MAAI,KAAK,UAAU,KAAM,QAAO,SAAS,WAAW,IAAI,CAAC;AACzD,QAAM,OAAc,CAAC;AACrB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,MAAM;AAC1C,eAAW,KAAK,WAAW,KAAK,SAAS,GAAG,IAAI,IAAI,CAAC,EAAG,MAAK,KAAK,CAAC;AACnE;AAAA,EACF;AACA,SAAO,SAAS,IAAI;AACtB;AAOA,SAAS,WAAW,MAAc,MAAsB,IAAsC;AAC5F,MAAI;AAGJ,QAAM,SAAiB,CAAC;AACxB,QAAM,SAAS,MAAO,OAAO,KAAK,SAAS;AAC3C,QAAM,OAAO,CAAC,MAAoB;AAChC,UAAM,IAAI,OAAQ,KAAK,CAAC,IAAe;AACvC,UAAM,IAAI,EAAE,GAAG,EAAE;AACjB,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,GAAG,KAAK,CAAC,EAAU,IAAI,IAAI,GAAG,GAAG,CAAC;AACxE,WAAO;AAAA,EACT;AACA,QAAM,QAAQ,IAAI,MAAM,QAAQ;AAAA,IAC9B,IAAI,GAAG,GAAG;AACR,UAAI,MAAM,SAAU,QAAO,OAAO;AAClC,YAAM,QAAI,uBAAW,GAAG,OAAO,CAAC;AAChC,aAAO,KAAK,IAAI,KAAK,CAAC,IAAI,EAAE,CAAiB;AAAA,IAC/C;AAAA,IACA,KAAK,CAAC,GAAG,UAAM,uBAAW,GAAG,OAAO,CAAC,KAAK,KAAK,KAAK;AAAA,EACtD,CAAC;AACD,SAAO;AAAA,IACL;AAAA;AAAA,IAEA,KAAK,OAAgC;AACnC,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,cACP,MACA,MACA,IACA,MACA,QACa;AAEb,QAAM,SAAS,IAAI,0BAAe;AAAA,IAChC;AAAA,IACA,GAAG;AAAA,IACH,eAAe;AAAA;AAAA;AAAA,IAGf,eAAe;AAAA,EACjB,CAAC;AAMD,QAAM,MAAM,OAAO;AACnB,MAAI,IAAI;AACR,QAAM,OAAO,CAAC,GAAW,MAAc;AACrC,UAAM,IAAI,GAAG,GAAG,CAAC;AACjB,QAAI,EAAE,SAAS,CAAC,KAAK,KAAK,KAAK,CAAC,EAAG,QAAO;AAC1C,QAAI,OAAO;AACX,QAAI,MAAM;AACV,UAAM,QAAQ,EAAE,MAAM,KAAK;AAC3B,eAAW,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,GAAG,KAAK,IAAI,OAAO;AAChD,YAAM,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,MAAM,IAAI,MAAM,IAAI,wBAAO,MAAM,GAAG,GAAG,KAAK,IAAI;AAChF,UAAI,IAAI,KAAK;AACX,eAAO;AACP,cAAM;AAAA,MACR;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACA,QAAM,OAAO,WAAW,MAAM,MAAM,IAAI;AACxC,SAAO,QAAQ,KAAK,KAAK;AACzB,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,SAAS,KAAK,KAAK;AACzB,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAE3C,YAAI;AACJ,aAAK,KAAK,KAAK;AACf,YAAI;AAEF,iBAAO,OAAO,OAAO,CAAC;AAAA,QACxB,UAAE;AACA,eAAK,KAAK,MAAS;AAAA,QACrB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAGA,SAAS,cACP,MACA,IACA,MACA,QACa;AACb,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,IAAI,KAAK,KAAK;AACpB,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAC3C,cAAM,OAAc,CAAC;AACrB,iBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,gBAAM,IAAI,MAAM,CAAC;AACjB,cAAI,OAAO;AACX,mBAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,kBAAM,IAAI,GAAG,GAAG,CAAC;AACjB,kBAAM,MAAM,MAAM,IAAI,IAAI,EAAE,WAAW,CAAC,IAAI,MAAM,EAAE,SAAS,CAAC,IAAI,OAAO;AACzE,kBAAM,IAAK,KAAK,CAAC,EAAU;AAC3B,gBAAI,MAAM,IAAI,KAAM,QAAO,MAAM;AAAA,UACnC;AACA,cAAI,OAAO,KAAK,QAAQ,OAAO,UAAW,MAAK,KAAK,EAAE,GAAG,GAAG,OAAO,KAAK,CAAC;AAAA,QAC3E;AACA,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAMO,SAAS,SAAS,MAA0B;AACjD,SAAO,YAAY;AAAA,IACjB,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC;AAAA,IAClD,CAAC,MAAM,EAAE;AAAA,EACX;AACF;","names":[]}
+{"version":3,"sources":["../../../src/features/search/search-index.ts"],"sourcesContent":["import { ExSearch } from 'ex-search'\nimport { getValue } from '../../column/get-value'\nimport '../../diagnostics/process-guard'\nimport type { Emit } from '../../diagnostics/types'\nimport type { RowModelContext } from '../../feature/types'\nimport { nullishOf, type SlicedFormatting } from '../../pipeline/column-cache'\nimport { resolveColumns } from '../../pipeline/columns'\nimport { arrayIndex } from '../../pipeline/row'\nimport { detectCapability, hasCapability } from '../../versioning/capability'\n\n/** Matcher settings (the client options of `search()`, defaults applied). */\nexport interface MatchConfig {\n  readonly threshold: number\n  readonly fuzzyMaxDistance: number\n  readonly caseSensitive: boolean\n}\n\n/**\n * One projected row as ex-search reads it: `i` is the row index (results map back by it, never by\n * identity: ex-search returns copies), `c0`, `c1`, … the formatted values of the searchable\n * columns. Flat keys also sidestep ex-search's top-level-key limitation for nested paths. Items\n * are built per chunk while matching and dropped after it (§9.3 memory): none is kept.\n */\nexport type Item = Record<string, string | number> & { readonly i: number }\n\n/** A match: the row index and its relevance (0–1, weight applied). */\nexport interface Hit {\n  readonly i: number\n  readonly score: number\n}\n\n/** The matcher of one (data, columns) pair: references to column values, no per-row copy. */\nexport interface SearchIndex {\n  /**\n   * A time-sliced job (§9.1) matching `query` (trimmed, non-blank) against the rows `rows` (row\n   * indices), in chunks: the hits in relevance order ({@link rankHits}).\n   */\n  match(rows: Uint32Array, query: string): Generator<undefined, Uint32Array, undefined>\n}\n\n/** Rows per chunk between two `yield`s of a scan (one value per row). */\nconst CHUNK = 4096\n\n/**\n * Cells (rows × searchable columns) matched per chunk between two `yield`s (#40): a table with\n * many searchable columns matches fewer rows per chunk, so a chunk stays a few milliseconds.\n */\nconst CELLS = 16_384\n\ninterface Key {\n  readonly name: string\n  readonly weight: number\n}\n\n/** The formatted value of row `r` of one searchable column. */\ntype Read = (r: number) => string\n\n// A throw (no `ExSearch`) counts as missing (hasCapability).\nconst hasExSearch = () =>\n  [ExSearch.prototype.search, ExSearch.prototype.setData, ExSearch.prototype.updateConfig].every(\n    (f) => typeof f === 'function',\n  )\n\n/** Feature-detects ex-search once (§8.5); `false` → the contains fallback. */\nexport function probeExSearch(emit: Emit): boolean {\n  // DK_DEP_INCOMPATIBLE is a development check (§8.5); the fallback is not.\n  return (process.env.NODE_ENV !== \"production\")\n    ? detectCapability(\n        hasExSearch,\n        {\n          feature: 'search',\n          dependency: 'ex-search',\n          capability: \"ExSearch (fuzzyMode: 'word')\",\n          supported: '^1.2.0',\n          fallback: 'case-insensitive exact / prefix / substring matching (no fuzzy matching)',\n        },\n        emit,\n      )\n    : hasCapability(hasExSearch)\n}\n\n/**\n * A numeric column's formatted value without cached strings (#27): what the default formatter\n * shows, `String(v)`, or `''` for a `null` / `undefined` cell (stored as `NaN`, like a real `NaN`,\n * which shows `'NaN'`).\n */\nfunction numberReader(values: Float64Array): Read {\n  const nullish = nullishOf(values)\n  return (r) => {\n    const v = values[r] as number\n    return !Number.isNaN(v) ? String(v) : nullish?.[r] ? '' : 'NaN'\n  }\n}\n\n/**\n * One pass over a column's values (`get(r)` per row, time-sliced by `CHUNK` rows): their text as\n * the default formatter shows a primitive (`String(v)`, `''` when nullish), read while matching\n * (#27: no formatted copy per row). `false` at the first object or function (a `Date`, an array):\n * the column cache formats those. `undefined` when a read throws. The values are read in NFC (#30),\n * normalised here, never per keystroke: the rows whose value is not NFC (rare: NFD input from\n * macOS, Thai or Vietnamese keyboards) keep their NFC form in a sparse map. Only values with a code\n * unit from U+0300 (combining marks and up) can change.\n */\nfunction* scan(\n  size: number,\n  get: (r: number) => unknown,\n): Generator<undefined, Read | false | undefined, undefined> {\n  const over = new Map<number, string>()\n  try {\n    for (let r = 0; r < size; r++) {\n      const v = get(r)\n      if (v !== null && (typeof v === 'object' || typeof v === 'function')) return false\n      if (typeof v === 'string' && /[\\u0300-\\uffff]/.test(v)) {\n        const n = v.normalize('NFC')\n        if (n !== v) over.set(r, n)\n      }\n      if (r % CHUNK === CHUNK - 1) yield\n    }\n  } catch {\n    return undefined\n  }\n  const read: Read = (r) => {\n    const v = get(r)\n    return v === null || v === undefined ? '' : String(v)\n  }\n  return over.size ? (r) => over.get(r) ?? read(r) : read\n}\n\n/**\n * Builds the matcher over the formatted values of the searchable columns (what the user sees), once\n * per data + columns, as a time-sliced job (§9.1): one column read per slice. Searchable columns are\n * those with a `search` weight > 0. When no column declares `search`, the columns of primitives\n * (strings, numbers, booleans, nullable) are searchable at weight 1, and a column of dates, arrays\n * or other objects is not (#27: it would keep a formatted string per row; it opts in with a\n * weight; `DK_SEARCH_COLUMN_SKIPPED` in development). `null` when nothing is searchable. Nothing\n * is copied per row (§9.3 memory): a key column of primitives without `format` is read from the\n * rows and stringified while matching (no values array, #27); an accessor column of primitives\n * reads the column cache's values the same way; only a column with `format` or an opted-in object\n * column (dates, arrays) keeps its formatted strings in the column cache. A key column whose read\n * throws goes through the column cache too, which reports it. The case is folded per chunk while\n * matching.\n */\nexport function* buildIndex(\n  ctx: RowModelContext,\n  config: MatchConfig,\n  useExSearch: boolean,\n): Generator<undefined, SearchIndex | null, undefined> {\n  const cols = resolveColumns(ctx.columns)\n  const declared = cols.list.some((c) => (c.def as { search?: unknown }).search !== undefined)\n  const data = ctx.data\n  const n = data.length\n  const keys: Key[] = []\n  const reads: Read[] = []\n  for (const col of cols.list) {\n    if (cols.byId.get(col.id) !== col) continue\n    const w = (col.def as { search?: unknown }).search\n    const weight = declared ? (typeof w === 'number' && w > 0 ? w : 0) : 1\n    if (weight === 0) continue\n    const path = col.format ? undefined : col.path\n    let failed = false\n    const onError = () => {\n      failed = true\n    }\n    let read = path ? yield* scan(n, (r) => getValue(data[r], path, onError)) : undefined\n    if (failed || (read === false && declared)) read = undefined\n    if (read === undefined) {\n      const values = ctx.getColumnValues(col.id)\n      if (!values) continue\n      yield // reading a column is a task of its own (#40)\n      const numbers = values instanceof Float64Array\n      if (numbers && !col.format) read = numberReader(values)\n      else {\n        // Zero-config: a column of objects is not searched, `format` or not (#27).\n        const own =\n          numbers || (declared && col.format) ? undefined : yield* scan(n, (r) => values[r])\n        if (own === false && !declared) read = false\n        else if (own && !col.format) read = own\n        else {\n          // Time-sliced in a table run (#40): dates or a heavy `format` are not one long task.\n          const sliced = (ctx as RowModelContext & SlicedFormatting).formattedSliced\n          const formatted = sliced ? yield* sliced(col.id) : ctx.getFormattedValues(col.id)\n          if (!formatted) continue\n          read = yield* scan(n, (r) => formatted[r])\n        }\n      }\n    }\n    if (!read) {\n      if ((process.env.NODE_ENV !== \"production\") && read === false) {\n        ctx.emit('DK_SEARCH_COLUMN_SKIPPED', { column: col.id, index: col.index })\n      }\n      continue\n    }\n    keys.push({ name: `c${keys.length}`, weight })\n    reads.push(read)\n    yield\n  }\n  if (keys.length === 0) return null\n  // Values (NFC already) are folded per chunk while matching; the query is normalised as well, so\n  // NFC and NFD spellings match each other like in the `contains` filter (#30).\n  const fold = (v: string) => (config.caseSensitive ? v : v.toLowerCase())\n  const foldQuery = (q: string) => fold(q.normalize('NFC'))\n  const at = (k: number, r: number) => fold((reads[k] as Read)(r))\n  return useExSearch\n    ? exSearchIndex(keys, at, foldQuery, config)\n    : containsIndex(keys, at, foldQuery, config)\n}\n\n/** Matches `rows` chunk by chunk (`score` decides alone, so chunks match like the whole). */\nfunction* chunked(\n  rows: Uint32Array,\n  columns: number,\n  matchChunk: (chunk: Uint32Array) => Hit[],\n): Generator<undefined, Uint32Array, undefined> {\n  const size = Math.max(256, Math.floor(CELLS / columns))\n  if (rows.length <= size) return rankHits(matchChunk(rows))\n  const hits: Hit[] = []\n  for (let k = 0; k < rows.length; k += size) {\n    for (const h of matchChunk(rows.subarray(k, k + size))) hits.push(h)\n    yield\n  }\n  return rankHits(hits)\n}\n\n/**\n * The projection as ex-search reads it: an array (`length`, index reads) of `{ i, c0, … }` items,\n * each built when read and not kept. It shows the chunk of rows being matched; ex-search holds\n * this one view (`setData` once).\n */\nfunction projection(keys: readonly Key[], at: (k: number, r: number) => string) {\n  let rows: Uint32Array = new Uint32Array(0)\n  // Never written, and its own `length` stays 0: `length` is read through the trap below (setting\n  // a large `length` would allocate a backing store).\n  const target: Item[] = []\n  const items = new Proxy(target, {\n    get(t, p) {\n      if (p === 'length') return rows.length\n      const j = arrayIndex(p, rows.length)\n      if (j < 0) return t[p as keyof Item[]]\n      const r = rows[j] as number\n      const o = { i: r } as Record<string, string | number>\n      for (let k = 0; k < keys.length; k++) o[(keys[k] as Key).name] = at(k, r)\n      return o\n    },\n    has: (t, p) => arrayIndex(p, rows.length) >= 0 || p in t,\n  })\n  return {\n    items,\n    /** Shows `chunk` (row indices). */\n    show(chunk: Uint32Array) {\n      rows = chunk\n    },\n  }\n}\n\nfunction exSearchIndex(\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  // Values and query are folded already (buildIndex).\n  const engine = new ExSearch<Item>({\n    keys: keys as Key[],\n    ...config,\n    caseSensitive: true,\n    // DataKable re-ranks with exact scores and an index tie-break; tim keeps ex-search's own pass\n    // stable and avoids radix's score quantisation.\n    sortAlgorithm: 'tim',\n  })\n  // A one-word query is measured against the whole value and each of its words, the best wins: a\n  // typo of one word of a multi-word cell matches (#10). ex-search's `fuzzyMode: 'word'` does it\n  // with its bounded matcher (#106). A query with whitespace is measured against whole values only,\n  // as before (a single word is never that query), and is decided once per query, not per row (#96).\n  // ex-search caches the words per position of its data: the view shows another chunk each time,\n  // and a cache kept over every row would hold a word list per row and column (tens of MB at 100k\n  // rows, §9.3), so it is dropped after each chunk (`updateConfig`) and values are split per chunk.\n  const view = projection(keys, at)\n  engine.setData(view.items)\n  return {\n    match(rows, query) {\n      const q = fold(query)\n      const fuzzyMode = /\\s/.test(q) ? 'value' : 'word'\n      return chunked(rows, keys.length, (chunk) => {\n        // Set per chunk, like the view: another job may have matched in between.\n        engine.updateConfig({ fuzzyMode })\n        view.show(chunk)\n        // Results are copies of the items (`Omit<Item, …>` drops `i` from the type, not the value).\n        const hits = engine.search(q) as unknown as Hit[]\n        engine.updateConfig({})\n        return hits\n      })\n    },\n  }\n}\n\n/** The fallback when ex-search is missing or incompatible: exact 1 · prefix 0.8 · substring 0.65. */\nfunction containsIndex(\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  return {\n    match(rows, query) {\n      const q = fold(query)\n      return chunked(rows, keys.length, (chunk) => {\n        const hits: Hit[] = []\n        for (let j = 0; j < chunk.length; j++) {\n          const r = chunk[j] as number\n          let best = 0\n          for (let k = 0; k < keys.length; k++) {\n            const v = at(k, r)\n            const raw = v === q ? 1 : v.startsWith(q) ? 0.8 : v.includes(q) ? 0.65 : 0\n            const w = (keys[k] as Key).weight\n            if (raw * w > best) best = raw * w\n          }\n          if (best > 0 && best >= config.threshold) hits.push({ i: r, score: best })\n        }\n        return hits\n      })\n    },\n  }\n}\n\n/**\n * Hits in relevance order (§6.3): score descending, ties by original row index. Enforced here so\n * the order never depends on the matcher's own ranking. Sorts `hits` (a fresh array) in place.\n */\nexport function rankHits(hits: Hit[]): Uint32Array {\n  return Uint32Array.from(\n    hits.sort((a, b) => b.score - a.score || a.i - b.i),\n    (h) => h.i,\n  )\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,uBAAyB;AACzB,uBAAyB;AACzB,2BAAO;AAGP,0BAAiD;AACjD,qBAA+B;AAC/B,iBAA2B;AAC3B,wBAAgD;AAiChD,IAAM,QAAQ;AAMd,IAAM,QAAQ;AAWd,IAAM,cAAc,MAClB,CAAC,0BAAS,UAAU,QAAQ,0BAAS,UAAU,SAAS,0BAAS,UAAU,YAAY,EAAE;AAAA,EACvF,CAAC,MAAM,OAAO,MAAM;AACtB;AAGK,SAAS,cAAc,MAAqB;AAEjD,SAAQ,QAAQ,IAAI,aAAa,mBAC7B;AAAA,IACE;AAAA,IACA;AAAA,MACE,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,MACX,UAAU;AAAA,IACZ;AAAA,IACA;AAAA,EACF,QACA,iCAAc,WAAW;AAC/B;AAOA,SAAS,aAAa,QAA4B;AAChD,QAAM,cAAU,+BAAU,MAAM;AAChC,SAAO,CAAC,MAAM;AACZ,UAAM,IAAI,OAAO,CAAC;AAClB,WAAO,CAAC,OAAO,MAAM,CAAC,IAAI,OAAO,CAAC,IAAI,UAAU,CAAC,IAAI,KAAK;AAAA,EAC5D;AACF;AAWA,UAAU,KACR,MACA,KAC2D;AAC3D,QAAM,OAAO,oBAAI,IAAoB;AACrC,MAAI;AACF,aAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,YAAM,IAAI,IAAI,CAAC;AACf,UAAI,MAAM,SAAS,OAAO,MAAM,YAAY,OAAO,MAAM,YAAa,QAAO;AAC7E,UAAI,OAAO,MAAM,YAAY,kBAAkB,KAAK,CAAC,GAAG;AACtD,cAAM,IAAI,EAAE,UAAU,KAAK;AAC3B,YAAI,MAAM,EAAG,MAAK,IAAI,GAAG,CAAC;AAAA,MAC5B;AACA,UAAI,IAAI,UAAU,QAAQ,EAAG;AAAA,IAC/B;AAAA,EACF,QAAQ;AACN,WAAO;AAAA,EACT;AACA,QAAM,OAAa,CAAC,MAAM;AACxB,UAAM,IAAI,IAAI,CAAC;AACf,WAAO,MAAM,QAAQ,MAAM,SAAY,KAAK,OAAO,CAAC;AAAA,EACtD;AACA,SAAO,KAAK,OAAO,CAAC,MAAM,KAAK,IAAI,CAAC,KAAK,KAAK,CAAC,IAAI;AACrD;AAgBO,UAAU,WACf,KACA,QACA,aACqD;AACrD,QAAM,WAAO,+BAAe,IAAI,OAAO;AACvC,QAAM,WAAW,KAAK,KAAK,KAAK,CAAC,MAAO,EAAE,IAA6B,WAAW,MAAS;AAC3F,QAAM,OAAO,IAAI;AACjB,QAAM,IAAI,KAAK;AACf,QAAM,OAAc,CAAC;AACrB,QAAM,QAAgB,CAAC;AACvB,aAAW,OAAO,KAAK,MAAM;AAC3B,QAAI,KAAK,KAAK,IAAI,IAAI,EAAE,MAAM,IAAK;AACnC,UAAM,IAAK,IAAI,IAA6B;AAC5C,UAAM,SAAS,WAAY,OAAO,MAAM,YAAY,IAAI,IAAI,IAAI,IAAK;AACrE,QAAI,WAAW,EAAG;AAClB,UAAM,OAAO,IAAI,SAAS,SAAY,IAAI;AAC1C,QAAI,SAAS;AACb,UAAM,UAAU,MAAM;AACpB,eAAS;AAAA,IACX;AACA,QAAI,OAAO,OAAO,OAAO,KAAK,GAAG,CAAC,UAAM,2BAAS,KAAK,CAAC,GAAG,MAAM,OAAO,CAAC,IAAI;AAC5E,QAAI,UAAW,SAAS,SAAS,SAAW,QAAO;AACnD,QAAI,SAAS,QAAW;AACtB,YAAM,SAAS,IAAI,gBAAgB,IAAI,EAAE;AACzC,UAAI,CAAC,OAAQ;AACb;AACA,YAAM,UAAU,kBAAkB;AAClC,UAAI,WAAW,CAAC,IAAI,OAAQ,QAAO,aAAa,MAAM;AAAA,WACjD;AAEH,cAAM,MACJ,WAAY,YAAY,IAAI,SAAU,SAAY,OAAO,KAAK,GAAG,CAAC,MAAM,OAAO,CAAC,CAAC;AACnF,YAAI,QAAQ,SAAS,CAAC,SAAU,QAAO;AAAA,iBAC9B,OAAO,CAAC,IAAI,OAAQ,QAAO;AAAA,aAC/B;AAEH,gBAAM,SAAU,IAA2C;AAC3D,gBAAM,YAAY,SAAS,OAAO,OAAO,IAAI,EAAE,IAAI,IAAI,mBAAmB,IAAI,EAAE;AAChF,cAAI,CAAC,UAAW;AAChB,iBAAO,OAAO,KAAK,GAAG,CAAC,MAAM,UAAU,CAAC,CAAC;AAAA,QAC3C;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM;AACT,UAAK,QAAQ,IAAI,aAAa,gBAAiB,SAAS,OAAO;AAC7D,YAAI,KAAK,4BAA4B,EAAE,QAAQ,IAAI,IAAI,OAAO,IAAI,MAAM,CAAC;AAAA,MAC3E;AACA;AAAA,IACF;AACA,SAAK,KAAK,EAAE,MAAM,IAAI,KAAK,MAAM,IAAI,OAAO,CAAC;AAC7C,UAAM,KAAK,IAAI;AACf;AAAA,EACF;AACA,MAAI,KAAK,WAAW,EAAG,QAAO;AAG9B,QAAM,OAAO,CAAC,MAAe,OAAO,gBAAgB,IAAI,EAAE,YAAY;AACtE,QAAM,YAAY,CAAC,MAAc,KAAK,EAAE,UAAU,KAAK,CAAC;AACxD,QAAM,KAAK,CAAC,GAAW,MAAc,KAAM,MAAM,CAAC,EAAW,CAAC,CAAC;AAC/D,SAAO,cACH,cAAc,MAAM,IAAI,WAAW,MAAM,IACzC,cAAc,MAAM,IAAI,WAAW,MAAM;AAC/C;AAGA,UAAU,QACR,MACA,SACA,YAC8C;AAC9C,QAAM,OAAO,KAAK,IAAI,KAAK,KAAK,MAAM,QAAQ,OAAO,CAAC;AACtD,MAAI,KAAK,UAAU,KAAM,QAAO,SAAS,WAAW,IAAI,CAAC;AACzD,QAAM,OAAc,CAAC;AACrB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,MAAM;AAC1C,eAAW,KAAK,WAAW,KAAK,SAAS,GAAG,IAAI,IAAI,CAAC,EAAG,MAAK,KAAK,CAAC;AACnE;AAAA,EACF;AACA,SAAO,SAAS,IAAI;AACtB;AAOA,SAAS,WAAW,MAAsB,IAAsC;AAC9E,MAAI,OAAoB,IAAI,YAAY,CAAC;AAGzC,QAAM,SAAiB,CAAC;AACxB,QAAM,QAAQ,IAAI,MAAM,QAAQ;AAAA,IAC9B,IAAI,GAAG,GAAG;AACR,UAAI,MAAM,SAAU,QAAO,KAAK;AAChC,YAAM,QAAI,uBAAW,GAAG,KAAK,MAAM;AACnC,UAAI,IAAI,EAAG,QAAO,EAAE,CAAiB;AACrC,YAAM,IAAI,KAAK,CAAC;AAChB,YAAM,IAAI,EAAE,GAAG,EAAE;AACjB,eAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,GAAG,KAAK,CAAC,EAAU,IAAI,IAAI,GAAG,GAAG,CAAC;AACxE,aAAO;AAAA,IACT;AAAA,IACA,KAAK,CAAC,GAAG,UAAM,uBAAW,GAAG,KAAK,MAAM,KAAK,KAAK,KAAK;AAAA,EACzD,CAAC;AACD,SAAO;AAAA,IACL;AAAA;AAAA,IAEA,KAAK,OAAoB;AACvB,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,cACP,MACA,IACA,MACA,QACa;AAEb,QAAM,SAAS,IAAI,0BAAe;AAAA,IAChC;AAAA,IACA,GAAG;AAAA,IACH,eAAe;AAAA;AAAA;AAAA,IAGf,eAAe;AAAA,EACjB,CAAC;AAQD,QAAM,OAAO,WAAW,MAAM,EAAE;AAChC,SAAO,QAAQ,KAAK,KAAK;AACzB,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,IAAI,KAAK,KAAK;AACpB,YAAM,YAAY,KAAK,KAAK,CAAC,IAAI,UAAU;AAC3C,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAE3C,eAAO,aAAa,EAAE,UAAU,CAAC;AACjC,aAAK,KAAK,KAAK;AAEf,cAAM,OAAO,OAAO,OAAO,CAAC;AAC5B,eAAO,aAAa,CAAC,CAAC;AACtB,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAGA,SAAS,cACP,MACA,IACA,MACA,QACa;AACb,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,IAAI,KAAK,KAAK;AACpB,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAC3C,cAAM,OAAc,CAAC;AACrB,iBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,gBAAM,IAAI,MAAM,CAAC;AACjB,cAAI,OAAO;AACX,mBAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,kBAAM,IAAI,GAAG,GAAG,CAAC;AACjB,kBAAM,MAAM,MAAM,IAAI,IAAI,EAAE,WAAW,CAAC,IAAI,MAAM,EAAE,SAAS,CAAC,IAAI,OAAO;AACzE,kBAAM,IAAK,KAAK,CAAC,EAAU;AAC3B,gBAAI,MAAM,IAAI,KAAM,QAAO,MAAM;AAAA,UACnC;AACA,cAAI,OAAO,KAAK,QAAQ,OAAO,UAAW,MAAK,KAAK,EAAE,GAAG,GAAG,OAAO,KAAK,CAAC;AAAA,QAC3E;AACA,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAMO,SAAS,SAAS,MAA0B;AACjD,SAAO,YAAY;AAAA,IACjB,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC;AAAA,IAClD,CAAC,MAAM,EAAE;AAAA,EACX;AACF;","names":[]}

package/dist/features/search/search-index.js

@@ -1,5 +1,5 @@
 // src/features/search/search-index.ts
-import { ExSearch, Scorer } from "ex-search";
+import { ExSearch } from "ex-search";
 import { getValue } from "../../column/get-value.js";
 import "../../diagnostics/process-guard.js";
 import { nullishOf } from "../../pipeline/column-cache.js";
@@ -8,7 +8,7 @@ import { arrayIndex } from "../../pipeline/row.js";
 import { detectCapability, hasCapability } from "../../versioning/capability.js";
 var CHUNK = 4096;
 var CELLS = 16384;
-var hasExSearch = () => [ExSearch.prototype.search, ExSearch.prototype.setData, Scorer.fuzzy].every(
+var hasExSearch = () => [ExSearch.prototype.search, ExSearch.prototype.setData, ExSearch.prototype.updateConfig].every(
   (f) => typeof f === "function"
 );
 function probeExSearch(emit) {
@@ -17,8 +17,8 @@ function probeExSearch(emit) {
     {
       feature: "search",
       dependency: "ex-search",
-      capability: "ExSearch, Scorer.fuzzy",
-      supported: "^1.1.1",
+      capability: "ExSearch (fuzzyMode: 'word')",
+      supported: "^1.2.0",
       fallback: "case-insensitive exact / prefix / substring matching (no fuzzy matching)"
     },
     emit
@@ -103,7 +103,7 @@ function* buildIndex(ctx, config, useExSearch) {
   const fold = (v) => config.caseSensitive ? v : v.toLowerCase();
   const foldQuery = (q) => fold(q.normalize("NFC"));
   const at = (k, r) => fold(reads[k](r));
-  return useExSearch ? exSearchIndex(ctx.data.length, keys, at, foldQuery, config) : containsIndex(keys, at, foldQuery, config);
+  return useExSearch ? exSearchIndex(keys, at, foldQuery, config) : containsIndex(keys, at, foldQuery, config);
 }
 function* chunked(rows, columns, matchChunk) {
   const size = Math.max(256, Math.floor(CELLS / columns));
@@ -115,33 +115,30 @@ function* chunked(rows, columns, matchChunk) {
   }
   return rankHits(hits);
 }
-function projection(size, keys, at) {
-  let rows;
+function projection(keys, at) {
+  let rows = new Uint32Array(0);
   const target = [];
-  const length = () => rows ? rows.length : size;
-  const item = (j) => {
-    const r = rows ? rows[j] : j;
-    const o = { i: r };
-    for (let k = 0; k < keys.length; k++) o[keys[k].name] = at(k, r);
-    return o;
-  };
   const items = new Proxy(target, {
     get(t, p) {
-      if (p === "length") return length();
-      const j = arrayIndex(p, length());
-      return j >= 0 ? item(j) : t[p];
+      if (p === "length") return rows.length;
+      const j = arrayIndex(p, rows.length);
+      if (j < 0) return t[p];
+      const r = rows[j];
+      const o = { i: r };
+      for (let k = 0; k < keys.length; k++) o[keys[k].name] = at(k, r);
+      return o;
     },
-    has: (t, p) => arrayIndex(p, length()) >= 0 || p in t
+    has: (t, p) => arrayIndex(p, rows.length) >= 0 || p in t
   });
   return {
     items,
-    /** Shows `chunk` (row indices), or every row again (`undefined`). */
+    /** Shows `chunk` (row indices). */
     show(chunk) {
       rows = chunk;
     }
   };
 }
-function exSearchIndex(size, keys, at, fold, config) {
+function exSearchIndex(keys, at, fold, config) {
   const engine = new ExSearch({
     keys,
     ...config,
@@ -150,36 +147,18 @@ function exSearchIndex(size, keys, at, fold, config) {
     // stable and avoids radix's score quantisation.
     sortAlgorithm: "tim"
   });
-  const max = config.fuzzyMaxDistance;
-  let q = "";
-  const near = (k, r) => {
-    const v = at(k, r);
-    if (v.includes(q) || /\s/.test(q)) return v;
-    let best = "";
-    let top = 0;
-    const words = v.split(/\s+/);
-    for (const w of words[1] ? [v, ...words] : words) {
-      const s = Math.abs(w.length - q.length) > max ? 0 : Scorer.fuzzy(w, q, max, true);
-      if (s > top) {
-        best = w;
-        top = s;
-      }
-    }
-    return best;
-  };
-  const view = projection(size, keys, near);
+  const view = projection(keys, at);
   engine.setData(view.items);
   return {
     match(rows, query) {
-      const folded = fold(query);
+      const q = fold(query);
+      const fuzzyMode = /\s/.test(q) ? "value" : "word";
       return chunked(rows, keys.length, (chunk) => {
-        q = folded;
+        engine.updateConfig({ fuzzyMode });
         view.show(chunk);
-        try {
-          return engine.search(q);
-        } finally {
-          view.show(void 0);
-        }
+        const hits = engine.search(q);
+        engine.updateConfig({});
+        return hits;
       });
     }
   };

package/dist/features/search/search-index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../src/features/search/search-index.ts"],"sourcesContent":["import { ExSearch, Scorer } from 'ex-search'\nimport { getValue } from '../../column/get-value'\nimport '../../diagnostics/process-guard'\nimport type { Emit } from '../../diagnostics/types'\nimport type { RowModelContext } from '../../feature/types'\nimport { nullishOf, type SlicedFormatting } from '../../pipeline/column-cache'\nimport { resolveColumns } from '../../pipeline/columns'\nimport { arrayIndex } from '../../pipeline/row'\nimport { detectCapability, hasCapability } from '../../versioning/capability'\n\n/** Matcher settings (the client options of `search()`, defaults applied). */\nexport interface MatchConfig {\n  readonly threshold: number\n  readonly fuzzyMaxDistance: number\n  readonly caseSensitive: boolean\n}\n\n/**\n * One projected row as ex-search reads it: `i` is the row index (results map back by it, never by\n * identity: ex-search returns copies), `c0`, `c1`, … the formatted values of the searchable\n * columns. Flat keys also sidestep ex-search's top-level-key limitation for nested paths. Items\n * are built per chunk while matching and dropped after it (§9.3 memory): none is kept.\n */\nexport type Item = Record<string, string | number> & { readonly i: number }\n\n/** A match: the row index and its relevance (0–1, weight applied). */\nexport interface Hit {\n  readonly i: number\n  readonly score: number\n}\n\n/** The matcher of one (data, columns) pair: references to column values, no per-row copy. */\nexport interface SearchIndex {\n  /**\n   * A time-sliced job (§9.1) matching `query` (trimmed, non-blank) against the rows `rows` (row\n   * indices), in chunks: the hits in relevance order ({@link rankHits}).\n   */\n  match(rows: Uint32Array, query: string): Generator<undefined, Uint32Array, undefined>\n}\n\n/** Rows per chunk between two `yield`s of a scan (one value per row). */\nconst CHUNK = 4096\n\n/**\n * Cells (rows × searchable columns) matched per chunk between two `yield`s (#40): a table with\n * many searchable columns matches fewer rows per chunk, so a chunk stays a few milliseconds.\n */\nconst CELLS = 16_384\n\ninterface Key {\n  readonly name: string\n  readonly weight: number\n}\n\n/** The formatted value of row `r` of one searchable column. */\ntype Read = (r: number) => string\n\n// A throw (no `ExSearch`) counts as missing (hasCapability).\nconst hasExSearch = () =>\n  [ExSearch.prototype.search, ExSearch.prototype.setData, Scorer.fuzzy].every(\n    (f) => typeof f === 'function',\n  )\n\n/** Feature-detects ex-search once (§8.5); `false` → the contains fallback. */\nexport function probeExSearch(emit: Emit): boolean {\n  // DK_DEP_INCOMPATIBLE is a development check (§8.5); the fallback is not.\n  return (process.env.NODE_ENV !== \"production\")\n    ? detectCapability(\n        hasExSearch,\n        {\n          feature: 'search',\n          dependency: 'ex-search',\n          capability: 'ExSearch, Scorer.fuzzy',\n          supported: '^1.1.1',\n          fallback: 'case-insensitive exact / prefix / substring matching (no fuzzy matching)',\n        },\n        emit,\n      )\n    : hasCapability(hasExSearch)\n}\n\n/**\n * A numeric column's formatted value without cached strings (#27): what the default formatter\n * shows, `String(v)`, or `''` for a `null` / `undefined` cell (stored as `NaN`, like a real `NaN`,\n * which shows `'NaN'`).\n */\nfunction numberReader(values: Float64Array): Read {\n  const nullish = nullishOf(values)\n  return (r) => {\n    const v = values[r] as number\n    return !Number.isNaN(v) ? String(v) : nullish?.[r] ? '' : 'NaN'\n  }\n}\n\n/**\n * One pass over a column's values (`get(r)` per row, time-sliced by `CHUNK` rows): their text as\n * the default formatter shows a primitive (`String(v)`, `''` when nullish), read while matching\n * (#27: no formatted copy per row). `false` at the first object or function (a `Date`, an array):\n * the column cache formats those. `undefined` when a read throws. The values are read in NFC (#30),\n * normalised here, never per keystroke: the rows whose value is not NFC (rare: NFD input from\n * macOS, Thai or Vietnamese keyboards) keep their NFC form in a sparse map. Only values with a code\n * unit from U+0300 (combining marks and up) can change.\n */\nfunction* scan(\n  size: number,\n  get: (r: number) => unknown,\n): Generator<undefined, Read | false | undefined, undefined> {\n  const over = new Map<number, string>()\n  try {\n    for (let r = 0; r < size; r++) {\n      const v = get(r)\n      if (v !== null && (typeof v === 'object' || typeof v === 'function')) return false\n      if (typeof v === 'string' && /[\\u0300-\\uffff]/.test(v)) {\n        const n = v.normalize('NFC')\n        if (n !== v) over.set(r, n)\n      }\n      if (r % CHUNK === CHUNK - 1) yield\n    }\n  } catch {\n    return undefined\n  }\n  const read: Read = (r) => {\n    const v = get(r)\n    return v === null || v === undefined ? '' : String(v)\n  }\n  return over.size ? (r) => over.get(r) ?? read(r) : read\n}\n\n/**\n * Builds the matcher over the formatted values of the searchable columns (what the user sees), once\n * per data + columns, as a time-sliced job (§9.1): one column read per slice. Searchable columns are\n * those with a `search` weight > 0. When no column declares `search`, the columns of primitives\n * (strings, numbers, booleans, nullable) are searchable at weight 1, and a column of dates, arrays\n * or other objects is not (#27: it would keep a formatted string per row; it opts in with a\n * weight; `DK_SEARCH_COLUMN_SKIPPED` in development). `null` when nothing is searchable. Nothing\n * is copied per row (§9.3 memory): a key column of primitives without `format` is read from the\n * rows and stringified while matching (no values array, #27); an accessor column of primitives\n * reads the column cache's values the same way; only a column with `format` or an opted-in object\n * column (dates, arrays) keeps its formatted strings in the column cache. A key column whose read\n * throws goes through the column cache too, which reports it. The case is folded per chunk while\n * matching.\n */\nexport function* buildIndex(\n  ctx: RowModelContext,\n  config: MatchConfig,\n  useExSearch: boolean,\n): Generator<undefined, SearchIndex | null, undefined> {\n  const cols = resolveColumns(ctx.columns)\n  const declared = cols.list.some((c) => (c.def as { search?: unknown }).search !== undefined)\n  const data = ctx.data\n  const n = data.length\n  const keys: Key[] = []\n  const reads: Read[] = []\n  for (const col of cols.list) {\n    if (cols.byId.get(col.id) !== col) continue\n    const w = (col.def as { search?: unknown }).search\n    const weight = declared ? (typeof w === 'number' && w > 0 ? w : 0) : 1\n    if (weight === 0) continue\n    const path = col.format ? undefined : col.path\n    let failed = false\n    const onError = () => {\n      failed = true\n    }\n    let read = path ? yield* scan(n, (r) => getValue(data[r], path, onError)) : undefined\n    if (failed || (read === false && declared)) read = undefined\n    if (read === undefined) {\n      const values = ctx.getColumnValues(col.id)\n      if (!values) continue\n      yield // reading a column is a task of its own (#40)\n      const numbers = values instanceof Float64Array\n      if (numbers && !col.format) read = numberReader(values)\n      else {\n        // Zero-config: a column of objects is not searched, `format` or not (#27).\n        const own =\n          numbers || (declared && col.format) ? undefined : yield* scan(n, (r) => values[r])\n        if (own === false && !declared) read = false\n        else if (own && !col.format) read = own\n        else {\n          // Time-sliced in a table run (#40): dates or a heavy `format` are not one long task.\n          const sliced = (ctx as RowModelContext & SlicedFormatting).formattedSliced\n          const formatted = sliced ? yield* sliced(col.id) : ctx.getFormattedValues(col.id)\n          if (!formatted) continue\n          read = yield* scan(n, (r) => formatted[r])\n        }\n      }\n    }\n    if (!read) {\n      if ((process.env.NODE_ENV !== \"production\") && read === false) {\n        ctx.emit('DK_SEARCH_COLUMN_SKIPPED', { column: col.id, index: col.index })\n      }\n      continue\n    }\n    keys.push({ name: `c${keys.length}`, weight })\n    reads.push(read)\n    yield\n  }\n  if (keys.length === 0) return null\n  // Values (NFC already) are folded per chunk while matching; the query is normalised as well, so\n  // NFC and NFD spellings match each other like in the `contains` filter (#30).\n  const fold = (v: string) => (config.caseSensitive ? v : v.toLowerCase())\n  const foldQuery = (q: string) => fold(q.normalize('NFC'))\n  const at = (k: number, r: number) => fold((reads[k] as Read)(r))\n  return useExSearch\n    ? exSearchIndex(ctx.data.length, keys, at, foldQuery, config)\n    : containsIndex(keys, at, foldQuery, config)\n}\n\n/** Matches `rows` chunk by chunk (`score` decides alone, so chunks match like the whole). */\nfunction* chunked(\n  rows: Uint32Array,\n  columns: number,\n  matchChunk: (chunk: Uint32Array) => Hit[],\n): Generator<undefined, Uint32Array, undefined> {\n  const size = Math.max(256, Math.floor(CELLS / columns))\n  if (rows.length <= size) return rankHits(matchChunk(rows))\n  const hits: Hit[] = []\n  for (let k = 0; k < rows.length; k += size) {\n    for (const h of matchChunk(rows.subarray(k, k + size))) hits.push(h)\n    yield\n  }\n  return rankHits(hits)\n}\n\n/**\n * The projection as ex-search reads it: an array (iteration, `length`, index reads) of `{ i, c0, … }`\n * items, each built when read and not kept. It shows every row of `data` (`size`), or one chunk of\n * rows while that chunk is matched; ex-search holds this one view (`setData` once).\n */\nfunction projection(size: number, keys: readonly Key[], at: (k: number, r: number) => string) {\n  let rows: Uint32Array | undefined\n  // Never written, and its own `length` stays 0: `length` is read through the trap below (setting\n  // a large `length` would allocate a backing store).\n  const target: Item[] = []\n  const length = () => (rows ? rows.length : size)\n  const item = (j: number): Item => {\n    const r = rows ? (rows[j] as number) : j\n    const o = { i: r } as Record<string, string | number>\n    for (let k = 0; k < keys.length; k++) o[(keys[k] as Key).name] = at(k, r)\n    return o as Item\n  }\n  const items = new Proxy(target, {\n    get(t, p) {\n      if (p === 'length') return length()\n      const j = arrayIndex(p, length())\n      return j >= 0 ? item(j) : t[p as keyof Item[]]\n    },\n    has: (t, p) => arrayIndex(p, length()) >= 0 || p in t,\n  })\n  return {\n    items,\n    /** Shows `chunk` (row indices), or every row again (`undefined`). */\n    show(chunk: Uint32Array | undefined) {\n      rows = chunk\n    },\n  }\n}\n\nfunction exSearchIndex(\n  size: number,\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  // Values and query are folded already (buildIndex).\n  const engine = new ExSearch<Item>({\n    keys: keys as Key[],\n    ...config,\n    caseSensitive: true,\n    // DataKable re-ranks with exact scores and an index tie-break; tim keeps ex-search's own pass\n    // stable and avoids radix's score quantisation.\n    sortAlgorithm: 'tim',\n  })\n  // ex-search measures the edit distance to the whole value, so a typo of one word of a multi-word\n  // cell never matched (#10). For a one-word query, a value without it is shown as the best fuzzy\n  // match of itself and its words, or as '' (scores 0, as the value would). A value with the query\n  // stays whole for the exact / prefix / substring scores. Only strings within `max` of the query's\n  // length are measured (an edit distance is at least the length difference).\n  const max = config.fuzzyMaxDistance\n  let q = ''\n  const near = (k: number, r: number) => {\n    const v = at(k, r)\n    if (v.includes(q) || /\\s/.test(q)) return v\n    let best = ''\n    let top = 0\n    const words = v.split(/\\s+/)\n    for (const w of words[1] ? [v, ...words] : words) {\n      const s = Math.abs(w.length - q.length) > max ? 0 : Scorer.fuzzy(w, q, max, true)\n      if (s > top) {\n        best = w\n        top = s\n      }\n    }\n    return best\n  }\n  const view = projection(size, keys, near)\n  engine.setData(view.items)\n  return {\n    match(rows, query) {\n      const folded = fold(query)\n      return chunked(rows, keys.length, (chunk) => {\n        // Set per chunk, like the view: another job may have matched in between.\n        q = folded\n        view.show(chunk)\n        try {\n          // Results are copies of the items (`Omit<Item, …>` drops `i` from the type, not the value).\n          return engine.search(q) as unknown as Hit[]\n        } finally {\n          view.show(undefined)\n        }\n      })\n    },\n  }\n}\n\n/** The fallback when ex-search is missing or incompatible: exact 1 · prefix 0.8 · substring 0.65. */\nfunction containsIndex(\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  return {\n    match(rows, query) {\n      const q = fold(query)\n      return chunked(rows, keys.length, (chunk) => {\n        const hits: Hit[] = []\n        for (let j = 0; j < chunk.length; j++) {\n          const r = chunk[j] as number\n          let best = 0\n          for (let k = 0; k < keys.length; k++) {\n            const v = at(k, r)\n            const raw = v === q ? 1 : v.startsWith(q) ? 0.8 : v.includes(q) ? 0.65 : 0\n            const w = (keys[k] as Key).weight\n            if (raw * w > best) best = raw * w\n          }\n          if (best > 0 && best >= config.threshold) hits.push({ i: r, score: best })\n        }\n        return hits\n      })\n    },\n  }\n}\n\n/**\n * Hits in relevance order (§6.3): score descending, ties by original row index. Enforced here so\n * the order never depends on the matcher's own ranking. Sorts `hits` (a fresh array) in place.\n */\nexport function rankHits(hits: Hit[]): Uint32Array {\n  return Uint32Array.from(\n    hits.sort((a, b) => b.score - a.score || a.i - b.i),\n    (h) => h.i,\n  )\n}\n"],"mappings":";AAAA,SAAS,UAAU,cAAc;AACjC,SAAS,gBAAgB;AACzB,OAAO;AAGP,SAAS,iBAAwC;AACjD,SAAS,sBAAsB;AAC/B,SAAS,kBAAkB;AAC3B,SAAS,kBAAkB,qBAAqB;AAiChD,IAAM,QAAQ;AAMd,IAAM,QAAQ;AAWd,IAAM,cAAc,MAClB,CAAC,SAAS,UAAU,QAAQ,SAAS,UAAU,SAAS,OAAO,KAAK,EAAE;AAAA,EACpE,CAAC,MAAM,OAAO,MAAM;AACtB;AAGK,SAAS,cAAc,MAAqB;AAEjD,SAAQ,QAAQ,IAAI,aAAa,eAC7B;AAAA,IACE;AAAA,IACA;AAAA,MACE,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,MACX,UAAU;AAAA,IACZ;AAAA,IACA;AAAA,EACF,IACA,cAAc,WAAW;AAC/B;AAOA,SAAS,aAAa,QAA4B;AAChD,QAAM,UAAU,UAAU,MAAM;AAChC,SAAO,CAAC,MAAM;AACZ,UAAM,IAAI,OAAO,CAAC;AAClB,WAAO,CAAC,OAAO,MAAM,CAAC,IAAI,OAAO,CAAC,IAAI,UAAU,CAAC,IAAI,KAAK;AAAA,EAC5D;AACF;AAWA,UAAU,KACR,MACA,KAC2D;AAC3D,QAAM,OAAO,oBAAI,IAAoB;AACrC,MAAI;AACF,aAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,YAAM,IAAI,IAAI,CAAC;AACf,UAAI,MAAM,SAAS,OAAO,MAAM,YAAY,OAAO,MAAM,YAAa,QAAO;AAC7E,UAAI,OAAO,MAAM,YAAY,kBAAkB,KAAK,CAAC,GAAG;AACtD,cAAM,IAAI,EAAE,UAAU,KAAK;AAC3B,YAAI,MAAM,EAAG,MAAK,IAAI,GAAG,CAAC;AAAA,MAC5B;AACA,UAAI,IAAI,UAAU,QAAQ,EAAG;AAAA,IAC/B;AAAA,EACF,QAAQ;AACN,WAAO;AAAA,EACT;AACA,QAAM,OAAa,CAAC,MAAM;AACxB,UAAM,IAAI,IAAI,CAAC;AACf,WAAO,MAAM,QAAQ,MAAM,SAAY,KAAK,OAAO,CAAC;AAAA,EACtD;AACA,SAAO,KAAK,OAAO,CAAC,MAAM,KAAK,IAAI,CAAC,KAAK,KAAK,CAAC,IAAI;AACrD;AAgBO,UAAU,WACf,KACA,QACA,aACqD;AACrD,QAAM,OAAO,eAAe,IAAI,OAAO;AACvC,QAAM,WAAW,KAAK,KAAK,KAAK,CAAC,MAAO,EAAE,IAA6B,WAAW,MAAS;AAC3F,QAAM,OAAO,IAAI;AACjB,QAAM,IAAI,KAAK;AACf,QAAM,OAAc,CAAC;AACrB,QAAM,QAAgB,CAAC;AACvB,aAAW,OAAO,KAAK,MAAM;AAC3B,QAAI,KAAK,KAAK,IAAI,IAAI,EAAE,MAAM,IAAK;AACnC,UAAM,IAAK,IAAI,IAA6B;AAC5C,UAAM,SAAS,WAAY,OAAO,MAAM,YAAY,IAAI,IAAI,IAAI,IAAK;AACrE,QAAI,WAAW,EAAG;AAClB,UAAM,OAAO,IAAI,SAAS,SAAY,IAAI;AAC1C,QAAI,SAAS;AACb,UAAM,UAAU,MAAM;AACpB,eAAS;AAAA,IACX;AACA,QAAI,OAAO,OAAO,OAAO,KAAK,GAAG,CAAC,MAAM,SAAS,KAAK,CAAC,GAAG,MAAM,OAAO,CAAC,IAAI;AAC5E,QAAI,UAAW,SAAS,SAAS,SAAW,QAAO;AACnD,QAAI,SAAS,QAAW;AACtB,YAAM,SAAS,IAAI,gBAAgB,IAAI,EAAE;AACzC,UAAI,CAAC,OAAQ;AACb;AACA,YAAM,UAAU,kBAAkB;AAClC,UAAI,WAAW,CAAC,IAAI,OAAQ,QAAO,aAAa,MAAM;AAAA,WACjD;AAEH,cAAM,MACJ,WAAY,YAAY,IAAI,SAAU,SAAY,OAAO,KAAK,GAAG,CAAC,MAAM,OAAO,CAAC,CAAC;AACnF,YAAI,QAAQ,SAAS,CAAC,SAAU,QAAO;AAAA,iBAC9B,OAAO,CAAC,IAAI,OAAQ,QAAO;AAAA,aAC/B;AAEH,gBAAM,SAAU,IAA2C;AAC3D,gBAAM,YAAY,SAAS,OAAO,OAAO,IAAI,EAAE,IAAI,IAAI,mBAAmB,IAAI,EAAE;AAChF,cAAI,CAAC,UAAW;AAChB,iBAAO,OAAO,KAAK,GAAG,CAAC,MAAM,UAAU,CAAC,CAAC;AAAA,QAC3C;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM;AACT,UAAK,QAAQ,IAAI,aAAa,gBAAiB,SAAS,OAAO;AAC7D,YAAI,KAAK,4BAA4B,EAAE,QAAQ,IAAI,IAAI,OAAO,IAAI,MAAM,CAAC;AAAA,MAC3E;AACA;AAAA,IACF;AACA,SAAK,KAAK,EAAE,MAAM,IAAI,KAAK,MAAM,IAAI,OAAO,CAAC;AAC7C,UAAM,KAAK,IAAI;AACf;AAAA,EACF;AACA,MAAI,KAAK,WAAW,EAAG,QAAO;AAG9B,QAAM,OAAO,CAAC,MAAe,OAAO,gBAAgB,IAAI,EAAE,YAAY;AACtE,QAAM,YAAY,CAAC,MAAc,KAAK,EAAE,UAAU,KAAK,CAAC;AACxD,QAAM,KAAK,CAAC,GAAW,MAAc,KAAM,MAAM,CAAC,EAAW,CAAC,CAAC;AAC/D,SAAO,cACH,cAAc,IAAI,KAAK,QAAQ,MAAM,IAAI,WAAW,MAAM,IAC1D,cAAc,MAAM,IAAI,WAAW,MAAM;AAC/C;AAGA,UAAU,QACR,MACA,SACA,YAC8C;AAC9C,QAAM,OAAO,KAAK,IAAI,KAAK,KAAK,MAAM,QAAQ,OAAO,CAAC;AACtD,MAAI,KAAK,UAAU,KAAM,QAAO,SAAS,WAAW,IAAI,CAAC;AACzD,QAAM,OAAc,CAAC;AACrB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,MAAM;AAC1C,eAAW,KAAK,WAAW,KAAK,SAAS,GAAG,IAAI,IAAI,CAAC,EAAG,MAAK,KAAK,CAAC;AACnE;AAAA,EACF;AACA,SAAO,SAAS,IAAI;AACtB;AAOA,SAAS,WAAW,MAAc,MAAsB,IAAsC;AAC5F,MAAI;AAGJ,QAAM,SAAiB,CAAC;AACxB,QAAM,SAAS,MAAO,OAAO,KAAK,SAAS;AAC3C,QAAM,OAAO,CAAC,MAAoB;AAChC,UAAM,IAAI,OAAQ,KAAK,CAAC,IAAe;AACvC,UAAM,IAAI,EAAE,GAAG,EAAE;AACjB,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,GAAG,KAAK,CAAC,EAAU,IAAI,IAAI,GAAG,GAAG,CAAC;AACxE,WAAO;AAAA,EACT;AACA,QAAM,QAAQ,IAAI,MAAM,QAAQ;AAAA,IAC9B,IAAI,GAAG,GAAG;AACR,UAAI,MAAM,SAAU,QAAO,OAAO;AAClC,YAAM,IAAI,WAAW,GAAG,OAAO,CAAC;AAChC,aAAO,KAAK,IAAI,KAAK,CAAC,IAAI,EAAE,CAAiB;AAAA,IAC/C;AAAA,IACA,KAAK,CAAC,GAAG,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,KAAK,KAAK;AAAA,EACtD,CAAC;AACD,SAAO;AAAA,IACL;AAAA;AAAA,IAEA,KAAK,OAAgC;AACnC,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,cACP,MACA,MACA,IACA,MACA,QACa;AAEb,QAAM,SAAS,IAAI,SAAe;AAAA,IAChC;AAAA,IACA,GAAG;AAAA,IACH,eAAe;AAAA;AAAA;AAAA,IAGf,eAAe;AAAA,EACjB,CAAC;AAMD,QAAM,MAAM,OAAO;AACnB,MAAI,IAAI;AACR,QAAM,OAAO,CAAC,GAAW,MAAc;AACrC,UAAM,IAAI,GAAG,GAAG,CAAC;AACjB,QAAI,EAAE,SAAS,CAAC,KAAK,KAAK,KAAK,CAAC,EAAG,QAAO;AAC1C,QAAI,OAAO;AACX,QAAI,MAAM;AACV,UAAM,QAAQ,EAAE,MAAM,KAAK;AAC3B,eAAW,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,GAAG,KAAK,IAAI,OAAO;AAChD,YAAM,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,MAAM,IAAI,MAAM,IAAI,OAAO,MAAM,GAAG,GAAG,KAAK,IAAI;AAChF,UAAI,IAAI,KAAK;AACX,eAAO;AACP,cAAM;AAAA,MACR;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACA,QAAM,OAAO,WAAW,MAAM,MAAM,IAAI;AACxC,SAAO,QAAQ,KAAK,KAAK;AACzB,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,SAAS,KAAK,KAAK;AACzB,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAE3C,YAAI;AACJ,aAAK,KAAK,KAAK;AACf,YAAI;AAEF,iBAAO,OAAO,OAAO,CAAC;AAAA,QACxB,UAAE;AACA,eAAK,KAAK,MAAS;AAAA,QACrB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAGA,SAAS,cACP,MACA,IACA,MACA,QACa;AACb,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,IAAI,KAAK,KAAK;AACpB,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAC3C,cAAM,OAAc,CAAC;AACrB,iBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,gBAAM,IAAI,MAAM,CAAC;AACjB,cAAI,OAAO;AACX,mBAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,kBAAM,IAAI,GAAG,GAAG,CAAC;AACjB,kBAAM,MAAM,MAAM,IAAI,IAAI,EAAE,WAAW,CAAC,IAAI,MAAM,EAAE,SAAS,CAAC,IAAI,OAAO;AACzE,kBAAM,IAAK,KAAK,CAAC,EAAU;AAC3B,gBAAI,MAAM,IAAI,KAAM,QAAO,MAAM;AAAA,UACnC;AACA,cAAI,OAAO,KAAK,QAAQ,OAAO,UAAW,MAAK,KAAK,EAAE,GAAG,GAAG,OAAO,KAAK,CAAC;AAAA,QAC3E;AACA,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAMO,SAAS,SAAS,MAA0B;AACjD,SAAO,YAAY;AAAA,IACjB,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC;AAAA,IAClD,CAAC,MAAM,EAAE;AAAA,EACX;AACF;","names":[]}
+{"version":3,"sources":["../../../src/features/search/search-index.ts"],"sourcesContent":["import { ExSearch } from 'ex-search'\nimport { getValue } from '../../column/get-value'\nimport '../../diagnostics/process-guard'\nimport type { Emit } from '../../diagnostics/types'\nimport type { RowModelContext } from '../../feature/types'\nimport { nullishOf, type SlicedFormatting } from '../../pipeline/column-cache'\nimport { resolveColumns } from '../../pipeline/columns'\nimport { arrayIndex } from '../../pipeline/row'\nimport { detectCapability, hasCapability } from '../../versioning/capability'\n\n/** Matcher settings (the client options of `search()`, defaults applied). */\nexport interface MatchConfig {\n  readonly threshold: number\n  readonly fuzzyMaxDistance: number\n  readonly caseSensitive: boolean\n}\n\n/**\n * One projected row as ex-search reads it: `i` is the row index (results map back by it, never by\n * identity: ex-search returns copies), `c0`, `c1`, … the formatted values of the searchable\n * columns. Flat keys also sidestep ex-search's top-level-key limitation for nested paths. Items\n * are built per chunk while matching and dropped after it (§9.3 memory): none is kept.\n */\nexport type Item = Record<string, string | number> & { readonly i: number }\n\n/** A match: the row index and its relevance (0–1, weight applied). */\nexport interface Hit {\n  readonly i: number\n  readonly score: number\n}\n\n/** The matcher of one (data, columns) pair: references to column values, no per-row copy. */\nexport interface SearchIndex {\n  /**\n   * A time-sliced job (§9.1) matching `query` (trimmed, non-blank) against the rows `rows` (row\n   * indices), in chunks: the hits in relevance order ({@link rankHits}).\n   */\n  match(rows: Uint32Array, query: string): Generator<undefined, Uint32Array, undefined>\n}\n\n/** Rows per chunk between two `yield`s of a scan (one value per row). */\nconst CHUNK = 4096\n\n/**\n * Cells (rows × searchable columns) matched per chunk between two `yield`s (#40): a table with\n * many searchable columns matches fewer rows per chunk, so a chunk stays a few milliseconds.\n */\nconst CELLS = 16_384\n\ninterface Key {\n  readonly name: string\n  readonly weight: number\n}\n\n/** The formatted value of row `r` of one searchable column. */\ntype Read = (r: number) => string\n\n// A throw (no `ExSearch`) counts as missing (hasCapability).\nconst hasExSearch = () =>\n  [ExSearch.prototype.search, ExSearch.prototype.setData, ExSearch.prototype.updateConfig].every(\n    (f) => typeof f === 'function',\n  )\n\n/** Feature-detects ex-search once (§8.5); `false` → the contains fallback. */\nexport function probeExSearch(emit: Emit): boolean {\n  // DK_DEP_INCOMPATIBLE is a development check (§8.5); the fallback is not.\n  return (process.env.NODE_ENV !== \"production\")\n    ? detectCapability(\n        hasExSearch,\n        {\n          feature: 'search',\n          dependency: 'ex-search',\n          capability: \"ExSearch (fuzzyMode: 'word')\",\n          supported: '^1.2.0',\n          fallback: 'case-insensitive exact / prefix / substring matching (no fuzzy matching)',\n        },\n        emit,\n      )\n    : hasCapability(hasExSearch)\n}\n\n/**\n * A numeric column's formatted value without cached strings (#27): what the default formatter\n * shows, `String(v)`, or `''` for a `null` / `undefined` cell (stored as `NaN`, like a real `NaN`,\n * which shows `'NaN'`).\n */\nfunction numberReader(values: Float64Array): Read {\n  const nullish = nullishOf(values)\n  return (r) => {\n    const v = values[r] as number\n    return !Number.isNaN(v) ? String(v) : nullish?.[r] ? '' : 'NaN'\n  }\n}\n\n/**\n * One pass over a column's values (`get(r)` per row, time-sliced by `CHUNK` rows): their text as\n * the default formatter shows a primitive (`String(v)`, `''` when nullish), read while matching\n * (#27: no formatted copy per row). `false` at the first object or function (a `Date`, an array):\n * the column cache formats those. `undefined` when a read throws. The values are read in NFC (#30),\n * normalised here, never per keystroke: the rows whose value is not NFC (rare: NFD input from\n * macOS, Thai or Vietnamese keyboards) keep their NFC form in a sparse map. Only values with a code\n * unit from U+0300 (combining marks and up) can change.\n */\nfunction* scan(\n  size: number,\n  get: (r: number) => unknown,\n): Generator<undefined, Read | false | undefined, undefined> {\n  const over = new Map<number, string>()\n  try {\n    for (let r = 0; r < size; r++) {\n      const v = get(r)\n      if (v !== null && (typeof v === 'object' || typeof v === 'function')) return false\n      if (typeof v === 'string' && /[\\u0300-\\uffff]/.test(v)) {\n        const n = v.normalize('NFC')\n        if (n !== v) over.set(r, n)\n      }\n      if (r % CHUNK === CHUNK - 1) yield\n    }\n  } catch {\n    return undefined\n  }\n  const read: Read = (r) => {\n    const v = get(r)\n    return v === null || v === undefined ? '' : String(v)\n  }\n  return over.size ? (r) => over.get(r) ?? read(r) : read\n}\n\n/**\n * Builds the matcher over the formatted values of the searchable columns (what the user sees), once\n * per data + columns, as a time-sliced job (§9.1): one column read per slice. Searchable columns are\n * those with a `search` weight > 0. When no column declares `search`, the columns of primitives\n * (strings, numbers, booleans, nullable) are searchable at weight 1, and a column of dates, arrays\n * or other objects is not (#27: it would keep a formatted string per row; it opts in with a\n * weight; `DK_SEARCH_COLUMN_SKIPPED` in development). `null` when nothing is searchable. Nothing\n * is copied per row (§9.3 memory): a key column of primitives without `format` is read from the\n * rows and stringified while matching (no values array, #27); an accessor column of primitives\n * reads the column cache's values the same way; only a column with `format` or an opted-in object\n * column (dates, arrays) keeps its formatted strings in the column cache. A key column whose read\n * throws goes through the column cache too, which reports it. The case is folded per chunk while\n * matching.\n */\nexport function* buildIndex(\n  ctx: RowModelContext,\n  config: MatchConfig,\n  useExSearch: boolean,\n): Generator<undefined, SearchIndex | null, undefined> {\n  const cols = resolveColumns(ctx.columns)\n  const declared = cols.list.some((c) => (c.def as { search?: unknown }).search !== undefined)\n  const data = ctx.data\n  const n = data.length\n  const keys: Key[] = []\n  const reads: Read[] = []\n  for (const col of cols.list) {\n    if (cols.byId.get(col.id) !== col) continue\n    const w = (col.def as { search?: unknown }).search\n    const weight = declared ? (typeof w === 'number' && w > 0 ? w : 0) : 1\n    if (weight === 0) continue\n    const path = col.format ? undefined : col.path\n    let failed = false\n    const onError = () => {\n      failed = true\n    }\n    let read = path ? yield* scan(n, (r) => getValue(data[r], path, onError)) : undefined\n    if (failed || (read === false && declared)) read = undefined\n    if (read === undefined) {\n      const values = ctx.getColumnValues(col.id)\n      if (!values) continue\n      yield // reading a column is a task of its own (#40)\n      const numbers = values instanceof Float64Array\n      if (numbers && !col.format) read = numberReader(values)\n      else {\n        // Zero-config: a column of objects is not searched, `format` or not (#27).\n        const own =\n          numbers || (declared && col.format) ? undefined : yield* scan(n, (r) => values[r])\n        if (own === false && !declared) read = false\n        else if (own && !col.format) read = own\n        else {\n          // Time-sliced in a table run (#40): dates or a heavy `format` are not one long task.\n          const sliced = (ctx as RowModelContext & SlicedFormatting).formattedSliced\n          const formatted = sliced ? yield* sliced(col.id) : ctx.getFormattedValues(col.id)\n          if (!formatted) continue\n          read = yield* scan(n, (r) => formatted[r])\n        }\n      }\n    }\n    if (!read) {\n      if ((process.env.NODE_ENV !== \"production\") && read === false) {\n        ctx.emit('DK_SEARCH_COLUMN_SKIPPED', { column: col.id, index: col.index })\n      }\n      continue\n    }\n    keys.push({ name: `c${keys.length}`, weight })\n    reads.push(read)\n    yield\n  }\n  if (keys.length === 0) return null\n  // Values (NFC already) are folded per chunk while matching; the query is normalised as well, so\n  // NFC and NFD spellings match each other like in the `contains` filter (#30).\n  const fold = (v: string) => (config.caseSensitive ? v : v.toLowerCase())\n  const foldQuery = (q: string) => fold(q.normalize('NFC'))\n  const at = (k: number, r: number) => fold((reads[k] as Read)(r))\n  return useExSearch\n    ? exSearchIndex(keys, at, foldQuery, config)\n    : containsIndex(keys, at, foldQuery, config)\n}\n\n/** Matches `rows` chunk by chunk (`score` decides alone, so chunks match like the whole). */\nfunction* chunked(\n  rows: Uint32Array,\n  columns: number,\n  matchChunk: (chunk: Uint32Array) => Hit[],\n): Generator<undefined, Uint32Array, undefined> {\n  const size = Math.max(256, Math.floor(CELLS / columns))\n  if (rows.length <= size) return rankHits(matchChunk(rows))\n  const hits: Hit[] = []\n  for (let k = 0; k < rows.length; k += size) {\n    for (const h of matchChunk(rows.subarray(k, k + size))) hits.push(h)\n    yield\n  }\n  return rankHits(hits)\n}\n\n/**\n * The projection as ex-search reads it: an array (`length`, index reads) of `{ i, c0, … }` items,\n * each built when read and not kept. It shows the chunk of rows being matched; ex-search holds\n * this one view (`setData` once).\n */\nfunction projection(keys: readonly Key[], at: (k: number, r: number) => string) {\n  let rows: Uint32Array = new Uint32Array(0)\n  // Never written, and its own `length` stays 0: `length` is read through the trap below (setting\n  // a large `length` would allocate a backing store).\n  const target: Item[] = []\n  const items = new Proxy(target, {\n    get(t, p) {\n      if (p === 'length') return rows.length\n      const j = arrayIndex(p, rows.length)\n      if (j < 0) return t[p as keyof Item[]]\n      const r = rows[j] as number\n      const o = { i: r } as Record<string, string | number>\n      for (let k = 0; k < keys.length; k++) o[(keys[k] as Key).name] = at(k, r)\n      return o\n    },\n    has: (t, p) => arrayIndex(p, rows.length) >= 0 || p in t,\n  })\n  return {\n    items,\n    /** Shows `chunk` (row indices). */\n    show(chunk: Uint32Array) {\n      rows = chunk\n    },\n  }\n}\n\nfunction exSearchIndex(\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  // Values and query are folded already (buildIndex).\n  const engine = new ExSearch<Item>({\n    keys: keys as Key[],\n    ...config,\n    caseSensitive: true,\n    // DataKable re-ranks with exact scores and an index tie-break; tim keeps ex-search's own pass\n    // stable and avoids radix's score quantisation.\n    sortAlgorithm: 'tim',\n  })\n  // A one-word query is measured against the whole value and each of its words, the best wins: a\n  // typo of one word of a multi-word cell matches (#10). ex-search's `fuzzyMode: 'word'` does it\n  // with its bounded matcher (#106). A query with whitespace is measured against whole values only,\n  // as before (a single word is never that query), and is decided once per query, not per row (#96).\n  // ex-search caches the words per position of its data: the view shows another chunk each time,\n  // and a cache kept over every row would hold a word list per row and column (tens of MB at 100k\n  // rows, §9.3), so it is dropped after each chunk (`updateConfig`) and values are split per chunk.\n  const view = projection(keys, at)\n  engine.setData(view.items)\n  return {\n    match(rows, query) {\n      const q = fold(query)\n      const fuzzyMode = /\\s/.test(q) ? 'value' : 'word'\n      return chunked(rows, keys.length, (chunk) => {\n        // Set per chunk, like the view: another job may have matched in between.\n        engine.updateConfig({ fuzzyMode })\n        view.show(chunk)\n        // Results are copies of the items (`Omit<Item, …>` drops `i` from the type, not the value).\n        const hits = engine.search(q) as unknown as Hit[]\n        engine.updateConfig({})\n        return hits\n      })\n    },\n  }\n}\n\n/** The fallback when ex-search is missing or incompatible: exact 1 · prefix 0.8 · substring 0.65. */\nfunction containsIndex(\n  keys: readonly Key[],\n  at: (k: number, r: number) => string,\n  fold: (v: string) => string,\n  config: MatchConfig,\n): SearchIndex {\n  return {\n    match(rows, query) {\n      const q = fold(query)\n      return chunked(rows, keys.length, (chunk) => {\n        const hits: Hit[] = []\n        for (let j = 0; j < chunk.length; j++) {\n          const r = chunk[j] as number\n          let best = 0\n          for (let k = 0; k < keys.length; k++) {\n            const v = at(k, r)\n            const raw = v === q ? 1 : v.startsWith(q) ? 0.8 : v.includes(q) ? 0.65 : 0\n            const w = (keys[k] as Key).weight\n            if (raw * w > best) best = raw * w\n          }\n          if (best > 0 && best >= config.threshold) hits.push({ i: r, score: best })\n        }\n        return hits\n      })\n    },\n  }\n}\n\n/**\n * Hits in relevance order (§6.3): score descending, ties by original row index. Enforced here so\n * the order never depends on the matcher's own ranking. Sorts `hits` (a fresh array) in place.\n */\nexport function rankHits(hits: Hit[]): Uint32Array {\n  return Uint32Array.from(\n    hits.sort((a, b) => b.score - a.score || a.i - b.i),\n    (h) => h.i,\n  )\n}\n"],"mappings":";AAAA,SAAS,gBAAgB;AACzB,SAAS,gBAAgB;AACzB,OAAO;AAGP,SAAS,iBAAwC;AACjD,SAAS,sBAAsB;AAC/B,SAAS,kBAAkB;AAC3B,SAAS,kBAAkB,qBAAqB;AAiChD,IAAM,QAAQ;AAMd,IAAM,QAAQ;AAWd,IAAM,cAAc,MAClB,CAAC,SAAS,UAAU,QAAQ,SAAS,UAAU,SAAS,SAAS,UAAU,YAAY,EAAE;AAAA,EACvF,CAAC,MAAM,OAAO,MAAM;AACtB;AAGK,SAAS,cAAc,MAAqB;AAEjD,SAAQ,QAAQ,IAAI,aAAa,eAC7B;AAAA,IACE;AAAA,IACA;AAAA,MACE,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,MACX,UAAU;AAAA,IACZ;AAAA,IACA;AAAA,EACF,IACA,cAAc,WAAW;AAC/B;AAOA,SAAS,aAAa,QAA4B;AAChD,QAAM,UAAU,UAAU,MAAM;AAChC,SAAO,CAAC,MAAM;AACZ,UAAM,IAAI,OAAO,CAAC;AAClB,WAAO,CAAC,OAAO,MAAM,CAAC,IAAI,OAAO,CAAC,IAAI,UAAU,CAAC,IAAI,KAAK;AAAA,EAC5D;AACF;AAWA,UAAU,KACR,MACA,KAC2D;AAC3D,QAAM,OAAO,oBAAI,IAAoB;AACrC,MAAI;AACF,aAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,YAAM,IAAI,IAAI,CAAC;AACf,UAAI,MAAM,SAAS,OAAO,MAAM,YAAY,OAAO,MAAM,YAAa,QAAO;AAC7E,UAAI,OAAO,MAAM,YAAY,kBAAkB,KAAK,CAAC,GAAG;AACtD,cAAM,IAAI,EAAE,UAAU,KAAK;AAC3B,YAAI,MAAM,EAAG,MAAK,IAAI,GAAG,CAAC;AAAA,MAC5B;AACA,UAAI,IAAI,UAAU,QAAQ,EAAG;AAAA,IAC/B;AAAA,EACF,QAAQ;AACN,WAAO;AAAA,EACT;AACA,QAAM,OAAa,CAAC,MAAM;AACxB,UAAM,IAAI,IAAI,CAAC;AACf,WAAO,MAAM,QAAQ,MAAM,SAAY,KAAK,OAAO,CAAC;AAAA,EACtD;AACA,SAAO,KAAK,OAAO,CAAC,MAAM,KAAK,IAAI,CAAC,KAAK,KAAK,CAAC,IAAI;AACrD;AAgBO,UAAU,WACf,KACA,QACA,aACqD;AACrD,QAAM,OAAO,eAAe,IAAI,OAAO;AACvC,QAAM,WAAW,KAAK,KAAK,KAAK,CAAC,MAAO,EAAE,IAA6B,WAAW,MAAS;AAC3F,QAAM,OAAO,IAAI;AACjB,QAAM,IAAI,KAAK;AACf,QAAM,OAAc,CAAC;AACrB,QAAM,QAAgB,CAAC;AACvB,aAAW,OAAO,KAAK,MAAM;AAC3B,QAAI,KAAK,KAAK,IAAI,IAAI,EAAE,MAAM,IAAK;AACnC,UAAM,IAAK,IAAI,IAA6B;AAC5C,UAAM,SAAS,WAAY,OAAO,MAAM,YAAY,IAAI,IAAI,IAAI,IAAK;AACrE,QAAI,WAAW,EAAG;AAClB,UAAM,OAAO,IAAI,SAAS,SAAY,IAAI;AAC1C,QAAI,SAAS;AACb,UAAM,UAAU,MAAM;AACpB,eAAS;AAAA,IACX;AACA,QAAI,OAAO,OAAO,OAAO,KAAK,GAAG,CAAC,MAAM,SAAS,KAAK,CAAC,GAAG,MAAM,OAAO,CAAC,IAAI;AAC5E,QAAI,UAAW,SAAS,SAAS,SAAW,QAAO;AACnD,QAAI,SAAS,QAAW;AACtB,YAAM,SAAS,IAAI,gBAAgB,IAAI,EAAE;AACzC,UAAI,CAAC,OAAQ;AACb;AACA,YAAM,UAAU,kBAAkB;AAClC,UAAI,WAAW,CAAC,IAAI,OAAQ,QAAO,aAAa,MAAM;AAAA,WACjD;AAEH,cAAM,MACJ,WAAY,YAAY,IAAI,SAAU,SAAY,OAAO,KAAK,GAAG,CAAC,MAAM,OAAO,CAAC,CAAC;AACnF,YAAI,QAAQ,SAAS,CAAC,SAAU,QAAO;AAAA,iBAC9B,OAAO,CAAC,IAAI,OAAQ,QAAO;AAAA,aAC/B;AAEH,gBAAM,SAAU,IAA2C;AAC3D,gBAAM,YAAY,SAAS,OAAO,OAAO,IAAI,EAAE,IAAI,IAAI,mBAAmB,IAAI,EAAE;AAChF,cAAI,CAAC,UAAW;AAChB,iBAAO,OAAO,KAAK,GAAG,CAAC,MAAM,UAAU,CAAC,CAAC;AAAA,QAC3C;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM;AACT,UAAK,QAAQ,IAAI,aAAa,gBAAiB,SAAS,OAAO;AAC7D,YAAI,KAAK,4BAA4B,EAAE,QAAQ,IAAI,IAAI,OAAO,IAAI,MAAM,CAAC;AAAA,MAC3E;AACA;AAAA,IACF;AACA,SAAK,KAAK,EAAE,MAAM,IAAI,KAAK,MAAM,IAAI,OAAO,CAAC;AAC7C,UAAM,KAAK,IAAI;AACf;AAAA,EACF;AACA,MAAI,KAAK,WAAW,EAAG,QAAO;AAG9B,QAAM,OAAO,CAAC,MAAe,OAAO,gBAAgB,IAAI,EAAE,YAAY;AACtE,QAAM,YAAY,CAAC,MAAc,KAAK,EAAE,UAAU,KAAK,CAAC;AACxD,QAAM,KAAK,CAAC,GAAW,MAAc,KAAM,MAAM,CAAC,EAAW,CAAC,CAAC;AAC/D,SAAO,cACH,cAAc,MAAM,IAAI,WAAW,MAAM,IACzC,cAAc,MAAM,IAAI,WAAW,MAAM;AAC/C;AAGA,UAAU,QACR,MACA,SACA,YAC8C;AAC9C,QAAM,OAAO,KAAK,IAAI,KAAK,KAAK,MAAM,QAAQ,OAAO,CAAC;AACtD,MAAI,KAAK,UAAU,KAAM,QAAO,SAAS,WAAW,IAAI,CAAC;AACzD,QAAM,OAAc,CAAC;AACrB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,MAAM;AAC1C,eAAW,KAAK,WAAW,KAAK,SAAS,GAAG,IAAI,IAAI,CAAC,EAAG,MAAK,KAAK,CAAC;AACnE;AAAA,EACF;AACA,SAAO,SAAS,IAAI;AACtB;AAOA,SAAS,WAAW,MAAsB,IAAsC;AAC9E,MAAI,OAAoB,IAAI,YAAY,CAAC;AAGzC,QAAM,SAAiB,CAAC;AACxB,QAAM,QAAQ,IAAI,MAAM,QAAQ;AAAA,IAC9B,IAAI,GAAG,GAAG;AACR,UAAI,MAAM,SAAU,QAAO,KAAK;AAChC,YAAM,IAAI,WAAW,GAAG,KAAK,MAAM;AACnC,UAAI,IAAI,EAAG,QAAO,EAAE,CAAiB;AACrC,YAAM,IAAI,KAAK,CAAC;AAChB,YAAM,IAAI,EAAE,GAAG,EAAE;AACjB,eAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,GAAG,KAAK,CAAC,EAAU,IAAI,IAAI,GAAG,GAAG,CAAC;AACxE,aAAO;AAAA,IACT;AAAA,IACA,KAAK,CAAC,GAAG,MAAM,WAAW,GAAG,KAAK,MAAM,KAAK,KAAK,KAAK;AAAA,EACzD,CAAC;AACD,SAAO;AAAA,IACL;AAAA;AAAA,IAEA,KAAK,OAAoB;AACvB,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,cACP,MACA,IACA,MACA,QACa;AAEb,QAAM,SAAS,IAAI,SAAe;AAAA,IAChC;AAAA,IACA,GAAG;AAAA,IACH,eAAe;AAAA;AAAA;AAAA,IAGf,eAAe;AAAA,EACjB,CAAC;AAQD,QAAM,OAAO,WAAW,MAAM,EAAE;AAChC,SAAO,QAAQ,KAAK,KAAK;AACzB,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,IAAI,KAAK,KAAK;AACpB,YAAM,YAAY,KAAK,KAAK,CAAC,IAAI,UAAU;AAC3C,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAE3C,eAAO,aAAa,EAAE,UAAU,CAAC;AACjC,aAAK,KAAK,KAAK;AAEf,cAAM,OAAO,OAAO,OAAO,CAAC;AAC5B,eAAO,aAAa,CAAC,CAAC;AACtB,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAGA,SAAS,cACP,MACA,IACA,MACA,QACa;AACb,SAAO;AAAA,IACL,MAAM,MAAM,OAAO;AACjB,YAAM,IAAI,KAAK,KAAK;AACpB,aAAO,QAAQ,MAAM,KAAK,QAAQ,CAAC,UAAU;AAC3C,cAAM,OAAc,CAAC;AACrB,iBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,gBAAM,IAAI,MAAM,CAAC;AACjB,cAAI,OAAO;AACX,mBAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,kBAAM,IAAI,GAAG,GAAG,CAAC;AACjB,kBAAM,MAAM,MAAM,IAAI,IAAI,EAAE,WAAW,CAAC,IAAI,MAAM,EAAE,SAAS,CAAC,IAAI,OAAO;AACzE,kBAAM,IAAK,KAAK,CAAC,EAAU;AAC3B,gBAAI,MAAM,IAAI,KAAM,QAAO,MAAM;AAAA,UACnC;AACA,cAAI,OAAO,KAAK,QAAQ,OAAO,UAAW,MAAK,KAAK,EAAE,GAAG,GAAG,OAAO,KAAK,CAAC;AAAA,QAC3E;AACA,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAMO,SAAS,SAAS,MAA0B;AACjD,SAAO,YAAY;AAAA,IACjB,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC;AAAA,IAClD,CAAC,MAAM,EAAE;AAAA,EACX;AACF;","names":[]}

package/dist/features/sorting/ranks.cjs

@@ -25,85 +25,152 @@ __export(ranks_exports, {
 module.exports = __toCommonJS(ranks_exports);
 var import_compare = require("./compare.cjs");
 var CHUNK = 4096;
-var RUN = 512;
-function* sortPositions(n, cmp) {
+var RUN = 1024;
+function* sortPositions(n, cmp, run) {
   let a = new Uint32Array(n);
+  let ea = new Uint8Array(n);
   for (let i = 0; i < n; i++) a[i] = i;
-  for (let lo = 0; lo < n; lo += RUN) {
-    a.subarray(lo, Math.min(lo + RUN, n)).sort(cmp);
+  for (let lo = 0; lo < n; lo += run) {
+    const hi = Math.min(lo + run, n);
+    let zero = false;
+    a.subarray(lo, hi).sort((x, y) => {
+      const c = cmp(x, y);
+      if (c === 0) zero = true;
+      return c;
+    });
+    if (zero)
+      for (let k = lo + 1; k < hi; k++) ea[k] = cmp(a[k - 1], a[k]) ? 0 : 1;
     yield;
   }
   let b = new Uint32Array(n);
+  let eb = new Uint8Array(n);
   let ops = 0;
-  for (let width = RUN; width < n; width *= 2) {
+  for (let width = run; width < n; width *= 2) {
     for (let lo = 0; lo < n; lo += 2 * width) {
       const mid = Math.min(lo + width, n);
       const hi = Math.min(lo + 2 * width, n);
       let i = lo;
       let j = mid;
-      let k = lo;
-      while (i < mid && j < hi) {
-        const x = a[i];
-        const y = a[j];
-        if (cmp(y, x) < 0) {
-          b[k++] = y;
-          j++;
-        } else {
-          b[k++] = x;
-          i++;
-        }
-        if (++ops === CHUNK) {
-          ops = 0;
-          yield;
-        }
+      let prev = -1;
+      let same = 0;
+      for (let k = lo; k < hi; k++) {
+        let c = 1;
+        let src;
+        if (i < mid && j < hi) {
+          c = cmp(a[j], a[i]);
+          src = c < 0 ? 1 : 0;
+          if (++ops === CHUNK) {
+            ops = 0;
+            yield;
+          }
+        } else src = i < mid ? 0 : 1;
+        const p = src ? j++ : i++;
+        b[k] = a[p];
+        eb[k] = prev === src ? ea[p] : prev < 0 ? 0 : same;
+        same = src === 0 && c === 0 ? 1 : 0;
+        prev = src;
       }
-      while (i < mid) b[k++] = a[i++];
-      while (j < hi) b[k++] = a[j++];
     }
     const t = a;
     a = b;
     b = t;
+    const e = ea;
+    ea = eb;
+    eb = e;
   }
-  return a;
+  return { order: a, eq: ea };
 }
-function* collationRanks(values, collator) {
+function* collationRanks(values, collator, prev = { values: [], ranks: [], reps: [] }) {
   const n = values.length;
-  const ids = /* @__PURE__ */ new Map();
-  const distinct = [];
-  const slot = new Int32Array(n);
+  const { values: was, ranks: wasRank, reps } = prev;
+  const m = reps.length;
+  const out = new Float64Array(n);
+  const used = new Uint8Array(m);
+  const freshIds = /* @__PURE__ */ new Map();
+  const fresh = [];
+  const gap = [];
+  let rowOf;
+  let shift = 0;
+  let ops = 0;
   for (let r = 0; r < n; r++) {
     const v = values[r];
     if (typeof v === "string") {
-      let id = ids.get(v);
-      if (id === void 0) {
-        id = distinct.length;
-        ids.set(v, id);
-        distinct.push(v);
+      let k;
+      const q = r + shift;
+      if (was[q] === v) k = wasRank[q];
+      else {
+        const id = freshIds.get(v);
+        if (id !== void 0) k = m + id;
+        else {
+          let lo = 0;
+          let hi = m;
+          while (lo < hi) {
+            const mid = lo + hi >>> 1;
+            const c = collator.compare(v, reps[mid]);
+            ops++;
+            if (c === 0) {
+              k = mid;
+              break;
+            }
+            if (c < 0) hi = mid;
+            else lo = mid + 1;
+          }
+          if (k === void 0) {
+            k = m + fresh.length;
+            freshIds.set(v, fresh.length);
+            fresh.push(v);
+            gap.push(lo - 0.5);
+          } else {
+            if (!rowOf) {
+              rowOf = new Int32Array(m);
+              for (let p = 0; p < was.length; p++) {
+                const x = wasRank[p];
+                if (!Number.isNaN(x)) rowOf[x] = p;
+              }
+            }
+            shift = rowOf[k] - r;
+          }
+        }
       }
-      slot[r] = id;
-    } else if ((0, import_compare.typeRank)(v) === import_compare.RANK.missing) slot[r] = -1;
+      if (k < m) used[k] = 1;
+      out[r] = k;
+    } else if ((0, import_compare.typeRank)(v) === import_compare.RANK.missing) out[r] = Number.NaN;
     else return null;
-    if (r % CHUNK === CHUNK - 1) yield;
+    if (++ops >= CHUNK) {
+      ops = 0;
+      yield;
+    }
   }
-  const order = yield* sortPositions(
-    distinct.length,
-    (a, b) => collator.compare(distinct[a], distinct[b])
+  if (!fresh.length && !used.includes(0)) return { ranks: out, state: { values, ranks: out, reps } };
+  const f = fresh.length;
+  const { order, eq } = yield* sortPositions(
+    f,
+    // By gap; two new strings in the same gap (cold: all of them) go to the collator.
+    (a, b) => gap[a] - gap[b] || collator.compare(fresh[a], fresh[b]),
+    // Comparing numbers is far cheaper: longer native runs, fewer merge passes.
+    m ? 8 * RUN : RUN
   );
-  const rankOf = new Float64Array(distinct.length);
-  let rank = 0;
-  for (let k = 0; k < order.length; k++) {
-    const id = order[k];
-    const s = distinct[id];
-    if (k > 0 && collator.compare(distinct[order[k - 1]], s) !== 0) rank++;
-    rankOf[id] = rank;
-    if (k % CHUNK === CHUNK - 1) yield;
+  const remap = new Float64Array(m + f);
+  const kept = [];
+  let rank = -1;
+  let j = 0;
+  for (let g = 0; g <= m; g++) {
+    while (j < f && gap[order[j]] < g) {
+      const id = order[j];
+      if (!eq[j]) kept[++rank] = fresh[id];
+      remap[m + id] = rank;
+      j++;
+    }
+    if (g < m && used[g]) {
+      kept[++rank] = reps[g];
+      remap[g] = rank;
+    }
   }
-  const out = new Float64Array(n);
   for (let r = 0; r < n; r++) {
-    const id = slot[r];
-    out[r] = id < 0 ? Number.NaN : rankOf[id];
+    const k = out[r];
+    if (!Number.isNaN(k)) out[r] = remap[k];
   }
-  return out;
+  return { ranks: out, state: { values, ranks: out, reps: kept } };
 }
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {